Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/buffer.c
4 *
5 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
6 */
7
8/*
9 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10 *
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
13 *
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
16 *
17 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 *
19 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
20 */
21
22#include <linux/kernel.h>
23#include <linux/sched/signal.h>
24#include <linux/syscalls.h>
25#include <linux/fs.h>
26#include <linux/iomap.h>
27#include <linux/mm.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>
53
54#include "internal.h"
55
56static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
57static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
58 enum rw_hint hint, struct writeback_control *wbc);
59
60#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
61
62inline void touch_buffer(struct buffer_head *bh)
63{
64 trace_block_touch_buffer(bh);
65 folio_mark_accessed(bh->b_folio);
66}
67EXPORT_SYMBOL(touch_buffer);
68
69void __lock_buffer(struct buffer_head *bh)
70{
71 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
72}
73EXPORT_SYMBOL(__lock_buffer);
74
75void unlock_buffer(struct buffer_head *bh)
76{
77 clear_bit_unlock(BH_Lock, &bh->b_state);
78 smp_mb__after_atomic();
79 wake_up_bit(&bh->b_state, BH_Lock);
80}
81EXPORT_SYMBOL(unlock_buffer);
82
83/*
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.
87 */
88void buffer_check_dirty_writeback(struct folio *folio,
89 bool *dirty, bool *writeback)
90{
91 struct buffer_head *head, *bh;
92 *dirty = false;
93 *writeback = false;
94
95 BUG_ON(!folio_test_locked(folio));
96
97 head = folio_buffers(folio);
98 if (!head)
99 return;
100
101 if (folio_test_writeback(folio))
102 *writeback = true;
103
104 bh = head;
105 do {
106 if (buffer_locked(bh))
107 *writeback = true;
108
109 if (buffer_dirty(bh))
110 *dirty = true;
111
112 bh = bh->b_this_page;
113 } while (bh != head);
114}
115
116/*
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.
120 */
121void __wait_on_buffer(struct buffer_head * bh)
122{
123 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
124}
125EXPORT_SYMBOL(__wait_on_buffer);
126
127static void buffer_io_error(struct buffer_head *bh, char *msg)
128{
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);
133}
134
135/*
136 * End-of-IO handler helper function which does not touch the bh after
137 * unlocking it.
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
141 * itself.
142 */
143static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
144{
145 if (uptodate) {
146 set_buffer_uptodate(bh);
147 } else {
148 /* This happens, due to failed read-ahead attempts. */
149 clear_buffer_uptodate(bh);
150 }
151 unlock_buffer(bh);
152}
153
154/*
155 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
156 * unlock the buffer.
157 */
158void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
159{
160 __end_buffer_read_notouch(bh, uptodate);
161 put_bh(bh);
162}
163EXPORT_SYMBOL(end_buffer_read_sync);
164
165void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
166{
167 if (uptodate) {
168 set_buffer_uptodate(bh);
169 } else {
170 buffer_io_error(bh, ", lost sync page write");
171 mark_buffer_write_io_error(bh);
172 clear_buffer_uptodate(bh);
173 }
174 unlock_buffer(bh);
175 put_bh(bh);
176}
177EXPORT_SYMBOL(end_buffer_write_sync);
178
179/*
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
183 * i_private_lock.
184 *
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.
188 */
189static struct buffer_head *
190__find_get_block_slow(struct block_device *bdev, sector_t block)
191{
192 struct address_space *bd_mapping = bdev->bd_mapping;
193 const int blkbits = bd_mapping->host->i_blkbits;
194 struct buffer_head *ret = NULL;
195 pgoff_t index;
196 struct buffer_head *bh;
197 struct buffer_head *head;
198 struct folio *folio;
199 int all_mapped = 1;
200 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
201
202 index = ((loff_t)block << blkbits) / PAGE_SIZE;
203 folio = __filemap_get_folio(bd_mapping, index, FGP_ACCESSED, 0);
204 if (IS_ERR(folio))
205 goto out;
206
207 spin_lock(&bd_mapping->i_private_lock);
208 head = folio_buffers(folio);
209 if (!head)
210 goto out_unlock;
211 bh = head;
212 do {
213 if (!buffer_mapped(bh))
214 all_mapped = 0;
215 else if (bh->b_blocknr == block) {
216 ret = bh;
217 get_bh(bh);
218 goto out_unlock;
219 }
220 bh = bh->b_this_page;
221 } while (bh != head);
222
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
227 */
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 << blkbits);
237 }
238out_unlock:
239 spin_unlock(&bd_mapping->i_private_lock);
240 folio_put(folio);
241out:
242 return ret;
243}
244
245static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
246{
247 unsigned long flags;
248 struct buffer_head *first;
249 struct buffer_head *tmp;
250 struct folio *folio;
251 int folio_uptodate = 1;
252
253 BUG_ON(!buffer_async_read(bh));
254
255 folio = bh->b_folio;
256 if (uptodate) {
257 set_buffer_uptodate(bh);
258 } else {
259 clear_buffer_uptodate(bh);
260 buffer_io_error(bh, ", async page read");
261 }
262
263 /*
264 * Be _very_ careful from here on. Bad things can happen if
265 * two buffer heads end IO at almost the same time and both
266 * decide that the page is now completely done.
267 */
268 first = folio_buffers(folio);
269 spin_lock_irqsave(&first->b_uptodate_lock, flags);
270 clear_buffer_async_read(bh);
271 unlock_buffer(bh);
272 tmp = bh;
273 do {
274 if (!buffer_uptodate(tmp))
275 folio_uptodate = 0;
276 if (buffer_async_read(tmp)) {
277 BUG_ON(!buffer_locked(tmp));
278 goto still_busy;
279 }
280 tmp = tmp->b_this_page;
281 } while (tmp != bh);
282 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
283
284 folio_end_read(folio, folio_uptodate);
285 return;
286
287still_busy:
288 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
289 return;
290}
291
292struct postprocess_bh_ctx {
293 struct work_struct work;
294 struct buffer_head *bh;
295};
296
297static void verify_bh(struct work_struct *work)
298{
299 struct postprocess_bh_ctx *ctx =
300 container_of(work, struct postprocess_bh_ctx, work);
301 struct buffer_head *bh = ctx->bh;
302 bool valid;
303
304 valid = fsverity_verify_blocks(bh->b_folio, bh->b_size, bh_offset(bh));
305 end_buffer_async_read(bh, valid);
306 kfree(ctx);
307}
308
309static bool need_fsverity(struct buffer_head *bh)
310{
311 struct folio *folio = bh->b_folio;
312 struct inode *inode = folio->mapping->host;
313
314 return fsverity_active(inode) &&
315 /* needed by ext4 */
316 folio->index < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
317}
318
319static void decrypt_bh(struct work_struct *work)
320{
321 struct postprocess_bh_ctx *ctx =
322 container_of(work, struct postprocess_bh_ctx, work);
323 struct buffer_head *bh = ctx->bh;
324 int err;
325
326 err = fscrypt_decrypt_pagecache_blocks(bh->b_folio, bh->b_size,
327 bh_offset(bh));
328 if (err == 0 && need_fsverity(bh)) {
329 /*
330 * We use different work queues for decryption and for verity
331 * because verity may require reading metadata pages that need
332 * decryption, and we shouldn't recurse to the same workqueue.
333 */
334 INIT_WORK(&ctx->work, verify_bh);
335 fsverity_enqueue_verify_work(&ctx->work);
336 return;
337 }
338 end_buffer_async_read(bh, err == 0);
339 kfree(ctx);
340}
341
342/*
343 * I/O completion handler for block_read_full_folio() - pages
344 * which come unlocked at the end of I/O.
345 */
346static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
347{
348 struct inode *inode = bh->b_folio->mapping->host;
349 bool decrypt = fscrypt_inode_uses_fs_layer_crypto(inode);
350 bool verify = need_fsverity(bh);
351
352 /* Decrypt (with fscrypt) and/or verify (with fsverity) if needed. */
353 if (uptodate && (decrypt || verify)) {
354 struct postprocess_bh_ctx *ctx =
355 kmalloc(sizeof(*ctx), GFP_ATOMIC);
356
357 if (ctx) {
358 ctx->bh = bh;
359 if (decrypt) {
360 INIT_WORK(&ctx->work, decrypt_bh);
361 fscrypt_enqueue_decrypt_work(&ctx->work);
362 } else {
363 INIT_WORK(&ctx->work, verify_bh);
364 fsverity_enqueue_verify_work(&ctx->work);
365 }
366 return;
367 }
368 uptodate = 0;
369 }
370 end_buffer_async_read(bh, uptodate);
371}
372
373/*
374 * Completion handler for block_write_full_folio() - folios which are unlocked
375 * during I/O, and which have the writeback flag cleared upon I/O completion.
376 */
377static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
378{
379 unsigned long flags;
380 struct buffer_head *first;
381 struct buffer_head *tmp;
382 struct folio *folio;
383
384 BUG_ON(!buffer_async_write(bh));
385
386 folio = bh->b_folio;
387 if (uptodate) {
388 set_buffer_uptodate(bh);
389 } else {
390 buffer_io_error(bh, ", lost async page write");
391 mark_buffer_write_io_error(bh);
392 clear_buffer_uptodate(bh);
393 }
394
395 first = folio_buffers(folio);
396 spin_lock_irqsave(&first->b_uptodate_lock, flags);
397
398 clear_buffer_async_write(bh);
399 unlock_buffer(bh);
400 tmp = bh->b_this_page;
401 while (tmp != bh) {
402 if (buffer_async_write(tmp)) {
403 BUG_ON(!buffer_locked(tmp));
404 goto still_busy;
405 }
406 tmp = tmp->b_this_page;
407 }
408 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
409 folio_end_writeback(folio);
410 return;
411
412still_busy:
413 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
414 return;
415}
416
417/*
418 * If a page's buffers are under async readin (end_buffer_async_read
419 * completion) then there is a possibility that another thread of
420 * control could lock one of the buffers after it has completed
421 * but while some of the other buffers have not completed. This
422 * locked buffer would confuse end_buffer_async_read() into not unlocking
423 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
424 * that this buffer is not under async I/O.
425 *
426 * The page comes unlocked when it has no locked buffer_async buffers
427 * left.
428 *
429 * PageLocked prevents anyone starting new async I/O reads any of
430 * the buffers.
431 *
432 * PageWriteback is used to prevent simultaneous writeout of the same
433 * page.
434 *
435 * PageLocked prevents anyone from starting writeback of a page which is
436 * under read I/O (PageWriteback is only ever set against a locked page).
437 */
438static void mark_buffer_async_read(struct buffer_head *bh)
439{
440 bh->b_end_io = end_buffer_async_read_io;
441 set_buffer_async_read(bh);
442}
443
444static void mark_buffer_async_write_endio(struct buffer_head *bh,
445 bh_end_io_t *handler)
446{
447 bh->b_end_io = handler;
448 set_buffer_async_write(bh);
449}
450
451void mark_buffer_async_write(struct buffer_head *bh)
452{
453 mark_buffer_async_write_endio(bh, end_buffer_async_write);
454}
455EXPORT_SYMBOL(mark_buffer_async_write);
456
457
458/*
459 * fs/buffer.c contains helper functions for buffer-backed address space's
460 * fsync functions. A common requirement for buffer-based filesystems is
461 * that certain data from the backing blockdev needs to be written out for
462 * a successful fsync(). For example, ext2 indirect blocks need to be
463 * written back and waited upon before fsync() returns.
464 *
465 * The functions mark_buffer_dirty_inode(), fsync_inode_buffers(),
466 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
467 * management of a list of dependent buffers at ->i_mapping->i_private_list.
468 *
469 * Locking is a little subtle: try_to_free_buffers() will remove buffers
470 * from their controlling inode's queue when they are being freed. But
471 * try_to_free_buffers() will be operating against the *blockdev* mapping
472 * at the time, not against the S_ISREG file which depends on those buffers.
473 * So the locking for i_private_list is via the i_private_lock in the address_space
474 * which backs the buffers. Which is different from the address_space
475 * against which the buffers are listed. So for a particular address_space,
476 * mapping->i_private_lock does *not* protect mapping->i_private_list! In fact,
477 * mapping->i_private_list will always be protected by the backing blockdev's
478 * ->i_private_lock.
479 *
480 * Which introduces a requirement: all buffers on an address_space's
481 * ->i_private_list must be from the same address_space: the blockdev's.
482 *
483 * address_spaces which do not place buffers at ->i_private_list via these
484 * utility functions are free to use i_private_lock and i_private_list for
485 * whatever they want. The only requirement is that list_empty(i_private_list)
486 * be true at clear_inode() time.
487 *
488 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
489 * filesystems should do that. invalidate_inode_buffers() should just go
490 * BUG_ON(!list_empty).
491 *
492 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
493 * take an address_space, not an inode. And it should be called
494 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
495 * queued up.
496 *
497 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
498 * list if it is already on a list. Because if the buffer is on a list,
499 * it *must* already be on the right one. If not, the filesystem is being
500 * silly. This will save a ton of locking. But first we have to ensure
501 * that buffers are taken *off* the old inode's list when they are freed
502 * (presumably in truncate). That requires careful auditing of all
503 * filesystems (do it inside bforget()). It could also be done by bringing
504 * b_inode back.
505 */
506
507/*
508 * The buffer's backing address_space's i_private_lock must be held
509 */
510static void __remove_assoc_queue(struct buffer_head *bh)
511{
512 list_del_init(&bh->b_assoc_buffers);
513 WARN_ON(!bh->b_assoc_map);
514 bh->b_assoc_map = NULL;
515}
516
517int inode_has_buffers(struct inode *inode)
518{
519 return !list_empty(&inode->i_data.i_private_list);
520}
521
522/*
523 * osync is designed to support O_SYNC io. It waits synchronously for
524 * all already-submitted IO to complete, but does not queue any new
525 * writes to the disk.
526 *
527 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
528 * as you dirty the buffers, and then use osync_inode_buffers to wait for
529 * completion. Any other dirty buffers which are not yet queued for
530 * write will not be flushed to disk by the osync.
531 */
532static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
533{
534 struct buffer_head *bh;
535 struct list_head *p;
536 int err = 0;
537
538 spin_lock(lock);
539repeat:
540 list_for_each_prev(p, list) {
541 bh = BH_ENTRY(p);
542 if (buffer_locked(bh)) {
543 get_bh(bh);
544 spin_unlock(lock);
545 wait_on_buffer(bh);
546 if (!buffer_uptodate(bh))
547 err = -EIO;
548 brelse(bh);
549 spin_lock(lock);
550 goto repeat;
551 }
552 }
553 spin_unlock(lock);
554 return err;
555}
556
557/**
558 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
559 * @mapping: the mapping which wants those buffers written
560 *
561 * Starts I/O against the buffers at mapping->i_private_list, and waits upon
562 * that I/O.
563 *
564 * Basically, this is a convenience function for fsync().
565 * @mapping is a file or directory which needs those buffers to be written for
566 * a successful fsync().
567 */
568int sync_mapping_buffers(struct address_space *mapping)
569{
570 struct address_space *buffer_mapping = mapping->i_private_data;
571
572 if (buffer_mapping == NULL || list_empty(&mapping->i_private_list))
573 return 0;
574
575 return fsync_buffers_list(&buffer_mapping->i_private_lock,
576 &mapping->i_private_list);
577}
578EXPORT_SYMBOL(sync_mapping_buffers);
579
580/**
581 * generic_buffers_fsync_noflush - generic buffer fsync implementation
582 * for simple filesystems with no inode lock
583 *
584 * @file: file to synchronize
585 * @start: start offset in bytes
586 * @end: end offset in bytes (inclusive)
587 * @datasync: only synchronize essential metadata if true
588 *
589 * This is a generic implementation of the fsync method for simple
590 * filesystems which track all non-inode metadata in the buffers list
591 * hanging off the address_space structure.
592 */
593int generic_buffers_fsync_noflush(struct file *file, loff_t start, loff_t end,
594 bool datasync)
595{
596 struct inode *inode = file->f_mapping->host;
597 int err;
598 int ret;
599
600 err = file_write_and_wait_range(file, start, end);
601 if (err)
602 return err;
603
604 ret = sync_mapping_buffers(inode->i_mapping);
605 if (!(inode->i_state & I_DIRTY_ALL))
606 goto out;
607 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
608 goto out;
609
610 err = sync_inode_metadata(inode, 1);
611 if (ret == 0)
612 ret = err;
613
614out:
615 /* check and advance again to catch errors after syncing out buffers */
616 err = file_check_and_advance_wb_err(file);
617 if (ret == 0)
618 ret = err;
619 return ret;
620}
621EXPORT_SYMBOL(generic_buffers_fsync_noflush);
622
623/**
624 * generic_buffers_fsync - generic buffer fsync implementation
625 * for simple filesystems with no inode lock
626 *
627 * @file: file to synchronize
628 * @start: start offset in bytes
629 * @end: end offset in bytes (inclusive)
630 * @datasync: only synchronize essential metadata if true
631 *
632 * This is a generic implementation of the fsync method for simple
633 * filesystems which track all non-inode metadata in the buffers list
634 * hanging off the address_space structure. This also makes sure that
635 * a device cache flush operation is called at the end.
636 */
637int generic_buffers_fsync(struct file *file, loff_t start, loff_t end,
638 bool datasync)
639{
640 struct inode *inode = file->f_mapping->host;
641 int ret;
642
643 ret = generic_buffers_fsync_noflush(file, start, end, datasync);
644 if (!ret)
645 ret = blkdev_issue_flush(inode->i_sb->s_bdev);
646 return ret;
647}
648EXPORT_SYMBOL(generic_buffers_fsync);
649
650/*
651 * Called when we've recently written block `bblock', and it is known that
652 * `bblock' was for a buffer_boundary() buffer. This means that the block at
653 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
654 * dirty, schedule it for IO. So that indirects merge nicely with their data.
655 */
656void write_boundary_block(struct block_device *bdev,
657 sector_t bblock, unsigned blocksize)
658{
659 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
660 if (bh) {
661 if (buffer_dirty(bh))
662 write_dirty_buffer(bh, 0);
663 put_bh(bh);
664 }
665}
666
667void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
668{
669 struct address_space *mapping = inode->i_mapping;
670 struct address_space *buffer_mapping = bh->b_folio->mapping;
671
672 mark_buffer_dirty(bh);
673 if (!mapping->i_private_data) {
674 mapping->i_private_data = buffer_mapping;
675 } else {
676 BUG_ON(mapping->i_private_data != buffer_mapping);
677 }
678 if (!bh->b_assoc_map) {
679 spin_lock(&buffer_mapping->i_private_lock);
680 list_move_tail(&bh->b_assoc_buffers,
681 &mapping->i_private_list);
682 bh->b_assoc_map = mapping;
683 spin_unlock(&buffer_mapping->i_private_lock);
684 }
685}
686EXPORT_SYMBOL(mark_buffer_dirty_inode);
687
688/**
689 * block_dirty_folio - Mark a folio as dirty.
690 * @mapping: The address space containing this folio.
691 * @folio: The folio to mark dirty.
692 *
693 * Filesystems which use buffer_heads can use this function as their
694 * ->dirty_folio implementation. Some filesystems need to do a little
695 * work before calling this function. Filesystems which do not use
696 * buffer_heads should call filemap_dirty_folio() instead.
697 *
698 * If the folio has buffers, the uptodate buffers are set dirty, to
699 * preserve dirty-state coherency between the folio and the buffers.
700 * Buffers added to a dirty folio are created dirty.
701 *
702 * The buffers are dirtied before the folio is dirtied. There's a small
703 * race window in which writeback may see the folio cleanness but not the
704 * buffer dirtiness. That's fine. If this code were to set the folio
705 * dirty before the buffers, writeback could clear the folio dirty flag,
706 * see a bunch of clean buffers and we'd end up with dirty buffers/clean
707 * folio on the dirty folio list.
708 *
709 * We use i_private_lock to lock against try_to_free_buffers() while
710 * using the folio's buffer list. This also prevents clean buffers
711 * being added to the folio after it was set dirty.
712 *
713 * Context: May only be called from process context. Does not sleep.
714 * Caller must ensure that @folio cannot be truncated during this call,
715 * typically by holding the folio lock or having a page in the folio
716 * mapped and holding the page table lock.
717 *
718 * Return: True if the folio was dirtied; false if it was already dirtied.
719 */
720bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
721{
722 struct buffer_head *head;
723 bool newly_dirty;
724
725 spin_lock(&mapping->i_private_lock);
726 head = folio_buffers(folio);
727 if (head) {
728 struct buffer_head *bh = head;
729
730 do {
731 set_buffer_dirty(bh);
732 bh = bh->b_this_page;
733 } while (bh != head);
734 }
735 /*
736 * Lock out page's memcg migration to keep PageDirty
737 * synchronized with per-memcg dirty page counters.
738 */
739 newly_dirty = !folio_test_set_dirty(folio);
740 spin_unlock(&mapping->i_private_lock);
741
742 if (newly_dirty)
743 __folio_mark_dirty(folio, mapping, 1);
744
745 if (newly_dirty)
746 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
747
748 return newly_dirty;
749}
750EXPORT_SYMBOL(block_dirty_folio);
751
752/*
753 * Write out and wait upon a list of buffers.
754 *
755 * We have conflicting pressures: we want to make sure that all
756 * initially dirty buffers get waited on, but that any subsequently
757 * dirtied buffers don't. After all, we don't want fsync to last
758 * forever if somebody is actively writing to the file.
759 *
760 * Do this in two main stages: first we copy dirty buffers to a
761 * temporary inode list, queueing the writes as we go. Then we clean
762 * up, waiting for those writes to complete.
763 *
764 * During this second stage, any subsequent updates to the file may end
765 * up refiling the buffer on the original inode's dirty list again, so
766 * there is a chance we will end up with a buffer queued for write but
767 * not yet completed on that list. So, as a final cleanup we go through
768 * the osync code to catch these locked, dirty buffers without requeuing
769 * any newly dirty buffers for write.
770 */
771static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
772{
773 struct buffer_head *bh;
774 struct address_space *mapping;
775 int err = 0, err2;
776 struct blk_plug plug;
777 LIST_HEAD(tmp);
778
779 blk_start_plug(&plug);
780
781 spin_lock(lock);
782 while (!list_empty(list)) {
783 bh = BH_ENTRY(list->next);
784 mapping = bh->b_assoc_map;
785 __remove_assoc_queue(bh);
786 /* Avoid race with mark_buffer_dirty_inode() which does
787 * a lockless check and we rely on seeing the dirty bit */
788 smp_mb();
789 if (buffer_dirty(bh) || buffer_locked(bh)) {
790 list_add(&bh->b_assoc_buffers, &tmp);
791 bh->b_assoc_map = mapping;
792 if (buffer_dirty(bh)) {
793 get_bh(bh);
794 spin_unlock(lock);
795 /*
796 * Ensure any pending I/O completes so that
797 * write_dirty_buffer() actually writes the
798 * current contents - it is a noop if I/O is
799 * still in flight on potentially older
800 * contents.
801 */
802 write_dirty_buffer(bh, REQ_SYNC);
803
804 /*
805 * Kick off IO for the previous mapping. Note
806 * that we will not run the very last mapping,
807 * wait_on_buffer() will do that for us
808 * through sync_buffer().
809 */
810 brelse(bh);
811 spin_lock(lock);
812 }
813 }
814 }
815
816 spin_unlock(lock);
817 blk_finish_plug(&plug);
818 spin_lock(lock);
819
820 while (!list_empty(&tmp)) {
821 bh = BH_ENTRY(tmp.prev);
822 get_bh(bh);
823 mapping = bh->b_assoc_map;
824 __remove_assoc_queue(bh);
825 /* Avoid race with mark_buffer_dirty_inode() which does
826 * a lockless check and we rely on seeing the dirty bit */
827 smp_mb();
828 if (buffer_dirty(bh)) {
829 list_add(&bh->b_assoc_buffers,
830 &mapping->i_private_list);
831 bh->b_assoc_map = mapping;
832 }
833 spin_unlock(lock);
834 wait_on_buffer(bh);
835 if (!buffer_uptodate(bh))
836 err = -EIO;
837 brelse(bh);
838 spin_lock(lock);
839 }
840
841 spin_unlock(lock);
842 err2 = osync_buffers_list(lock, list);
843 if (err)
844 return err;
845 else
846 return err2;
847}
848
849/*
850 * Invalidate any and all dirty buffers on a given inode. We are
851 * probably unmounting the fs, but that doesn't mean we have already
852 * done a sync(). Just drop the buffers from the inode list.
853 *
854 * NOTE: we take the inode's blockdev's mapping's i_private_lock. Which
855 * assumes that all the buffers are against the blockdev.
856 */
857void invalidate_inode_buffers(struct inode *inode)
858{
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;
863
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);
868 }
869}
870EXPORT_SYMBOL(invalidate_inode_buffers);
871
872/*
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.
875 *
876 * Returns true if all buffers were removed.
877 */
878int remove_inode_buffers(struct inode *inode)
879{
880 int ret = 1;
881
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;
886
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)) {
891 ret = 0;
892 break;
893 }
894 __remove_assoc_queue(bh);
895 }
896 spin_unlock(&buffer_mapping->i_private_lock);
897 }
898 return ret;
899}
900
901/*
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
905 * buffers.
906 *
907 * The retry flag is used to differentiate async IO (paging, swapping)
908 * which may not fail from ordinary buffer allocations.
909 */
910struct buffer_head *folio_alloc_buffers(struct folio *folio, unsigned long size,
911 gfp_t gfp)
912{
913 struct buffer_head *bh, *head;
914 long offset;
915 struct mem_cgroup *memcg, *old_memcg;
916
917 /* The folio lock pins the memcg */
918 memcg = folio_memcg(folio);
919 old_memcg = set_active_memcg(memcg);
920
921 head = NULL;
922 offset = folio_size(folio);
923 while ((offset -= size) >= 0) {
924 bh = alloc_buffer_head(gfp);
925 if (!bh)
926 goto no_grow;
927
928 bh->b_this_page = head;
929 bh->b_blocknr = -1;
930 head = bh;
931
932 bh->b_size = size;
933
934 /* Link the buffer to its folio */
935 folio_set_bh(bh, folio, offset);
936 }
937out:
938 set_active_memcg(old_memcg);
939 return head;
940/*
941 * In case anything failed, we just free everything we got.
942 */
943no_grow:
944 if (head) {
945 do {
946 bh = head;
947 head = head->b_this_page;
948 free_buffer_head(bh);
949 } while (head);
950 }
951
952 goto out;
953}
954EXPORT_SYMBOL_GPL(folio_alloc_buffers);
955
956struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size)
957{
958 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
959
960 return folio_alloc_buffers(page_folio(page), size, gfp);
961}
962EXPORT_SYMBOL_GPL(alloc_page_buffers);
963
964static inline void link_dev_buffers(struct folio *folio,
965 struct buffer_head *head)
966{
967 struct buffer_head *bh, *tail;
968
969 bh = head;
970 do {
971 tail = bh;
972 bh = bh->b_this_page;
973 } while (bh);
974 tail->b_this_page = head;
975 folio_attach_private(folio, head);
976}
977
978static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
979{
980 sector_t retval = ~((sector_t)0);
981 loff_t sz = bdev_nr_bytes(bdev);
982
983 if (sz) {
984 unsigned int sizebits = blksize_bits(size);
985 retval = (sz >> sizebits);
986 }
987 return retval;
988}
989
990/*
991 * Initialise the state of a blockdev folio's buffers.
992 */
993static sector_t folio_init_buffers(struct folio *folio,
994 struct block_device *bdev, unsigned size)
995{
996 struct buffer_head *head = folio_buffers(folio);
997 struct buffer_head *bh = head;
998 bool uptodate = folio_test_uptodate(folio);
999 sector_t block = div_u64(folio_pos(folio), size);
1000 sector_t end_block = blkdev_max_block(bdev, size);
1001
1002 do {
1003 if (!buffer_mapped(bh)) {
1004 bh->b_end_io = NULL;
1005 bh->b_private = NULL;
1006 bh->b_bdev = bdev;
1007 bh->b_blocknr = block;
1008 if (uptodate)
1009 set_buffer_uptodate(bh);
1010 if (block < end_block)
1011 set_buffer_mapped(bh);
1012 }
1013 block++;
1014 bh = bh->b_this_page;
1015 } while (bh != head);
1016
1017 /*
1018 * Caller needs to validate requested block against end of device.
1019 */
1020 return end_block;
1021}
1022
1023/*
1024 * Create the page-cache folio that contains the requested block.
1025 *
1026 * This is used purely for blockdev mappings.
1027 *
1028 * Returns false if we have a failure which cannot be cured by retrying
1029 * without sleeping. Returns true if we succeeded, or the caller should retry.
1030 */
1031static bool grow_dev_folio(struct block_device *bdev, sector_t block,
1032 pgoff_t index, unsigned size, gfp_t gfp)
1033{
1034 struct address_space *mapping = bdev->bd_mapping;
1035 struct folio *folio;
1036 struct buffer_head *bh;
1037 sector_t end_block = 0;
1038
1039 folio = __filemap_get_folio(mapping, index,
1040 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1041 if (IS_ERR(folio))
1042 return false;
1043
1044 bh = folio_buffers(folio);
1045 if (bh) {
1046 if (bh->b_size == size) {
1047 end_block = folio_init_buffers(folio, bdev, size);
1048 goto unlock;
1049 }
1050
1051 /*
1052 * Retrying may succeed; for example the folio may finish
1053 * writeback, or buffers may be cleaned. This should not
1054 * happen very often; maybe we have old buffers attached to
1055 * this blockdev's page cache and we're trying to change
1056 * the block size?
1057 */
1058 if (!try_to_free_buffers(folio)) {
1059 end_block = ~0ULL;
1060 goto unlock;
1061 }
1062 }
1063
1064 bh = folio_alloc_buffers(folio, size, gfp | __GFP_ACCOUNT);
1065 if (!bh)
1066 goto unlock;
1067
1068 /*
1069 * Link the folio to the buffers and initialise them. Take the
1070 * lock to be atomic wrt __find_get_block(), which does not
1071 * run under the folio lock.
1072 */
1073 spin_lock(&mapping->i_private_lock);
1074 link_dev_buffers(folio, bh);
1075 end_block = folio_init_buffers(folio, bdev, size);
1076 spin_unlock(&mapping->i_private_lock);
1077unlock:
1078 folio_unlock(folio);
1079 folio_put(folio);
1080 return block < end_block;
1081}
1082
1083/*
1084 * Create buffers for the specified block device block's folio. If
1085 * that folio was dirty, the buffers are set dirty also. Returns false
1086 * if we've hit a permanent error.
1087 */
1088static bool grow_buffers(struct block_device *bdev, sector_t block,
1089 unsigned size, gfp_t gfp)
1090{
1091 loff_t pos;
1092
1093 /*
1094 * Check for a block which lies outside our maximum possible
1095 * pagecache index.
1096 */
1097 if (check_mul_overflow(block, (sector_t)size, &pos) || pos > MAX_LFS_FILESIZE) {
1098 printk(KERN_ERR "%s: requested out-of-range block %llu for device %pg\n",
1099 __func__, (unsigned long long)block,
1100 bdev);
1101 return false;
1102 }
1103
1104 /* Create a folio with the proper size buffers */
1105 return grow_dev_folio(bdev, block, pos / PAGE_SIZE, size, gfp);
1106}
1107
1108static struct buffer_head *
1109__getblk_slow(struct block_device *bdev, sector_t block,
1110 unsigned size, gfp_t gfp)
1111{
1112 /* Size must be multiple of hard sectorsize */
1113 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1114 (size < 512 || size > PAGE_SIZE))) {
1115 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1116 size);
1117 printk(KERN_ERR "logical block size: %d\n",
1118 bdev_logical_block_size(bdev));
1119
1120 dump_stack();
1121 return NULL;
1122 }
1123
1124 for (;;) {
1125 struct buffer_head *bh;
1126
1127 bh = __find_get_block(bdev, block, size);
1128 if (bh)
1129 return bh;
1130
1131 if (!grow_buffers(bdev, block, size, gfp))
1132 return NULL;
1133 }
1134}
1135
1136/*
1137 * The relationship between dirty buffers and dirty pages:
1138 *
1139 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1140 * the page is tagged dirty in the page cache.
1141 *
1142 * At all times, the dirtiness of the buffers represents the dirtiness of
1143 * subsections of the page. If the page has buffers, the page dirty bit is
1144 * merely a hint about the true dirty state.
1145 *
1146 * When a page is set dirty in its entirety, all its buffers are marked dirty
1147 * (if the page has buffers).
1148 *
1149 * When a buffer is marked dirty, its page is dirtied, but the page's other
1150 * buffers are not.
1151 *
1152 * Also. When blockdev buffers are explicitly read with bread(), they
1153 * individually become uptodate. But their backing page remains not
1154 * uptodate - even if all of its buffers are uptodate. A subsequent
1155 * block_read_full_folio() against that folio will discover all the uptodate
1156 * buffers, will set the folio uptodate and will perform no I/O.
1157 */
1158
1159/**
1160 * mark_buffer_dirty - mark a buffer_head as needing writeout
1161 * @bh: the buffer_head to mark dirty
1162 *
1163 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1164 * its backing page dirty, then tag the page as dirty in the page cache
1165 * and then attach the address_space's inode to its superblock's dirty
1166 * inode list.
1167 *
1168 * mark_buffer_dirty() is atomic. It takes bh->b_folio->mapping->i_private_lock,
1169 * i_pages lock and mapping->host->i_lock.
1170 */
1171void mark_buffer_dirty(struct buffer_head *bh)
1172{
1173 WARN_ON_ONCE(!buffer_uptodate(bh));
1174
1175 trace_block_dirty_buffer(bh);
1176
1177 /*
1178 * Very *carefully* optimize the it-is-already-dirty case.
1179 *
1180 * Don't let the final "is it dirty" escape to before we
1181 * perhaps modified the buffer.
1182 */
1183 if (buffer_dirty(bh)) {
1184 smp_mb();
1185 if (buffer_dirty(bh))
1186 return;
1187 }
1188
1189 if (!test_set_buffer_dirty(bh)) {
1190 struct folio *folio = bh->b_folio;
1191 struct address_space *mapping = NULL;
1192
1193 if (!folio_test_set_dirty(folio)) {
1194 mapping = folio->mapping;
1195 if (mapping)
1196 __folio_mark_dirty(folio, mapping, 0);
1197 }
1198 if (mapping)
1199 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1200 }
1201}
1202EXPORT_SYMBOL(mark_buffer_dirty);
1203
1204void mark_buffer_write_io_error(struct buffer_head *bh)
1205{
1206 set_buffer_write_io_error(bh);
1207 /* FIXME: do we need to set this in both places? */
1208 if (bh->b_folio && bh->b_folio->mapping)
1209 mapping_set_error(bh->b_folio->mapping, -EIO);
1210 if (bh->b_assoc_map) {
1211 mapping_set_error(bh->b_assoc_map, -EIO);
1212 errseq_set(&bh->b_assoc_map->host->i_sb->s_wb_err, -EIO);
1213 }
1214}
1215EXPORT_SYMBOL(mark_buffer_write_io_error);
1216
1217/**
1218 * __brelse - Release a buffer.
1219 * @bh: The buffer to release.
1220 *
1221 * This variant of brelse() can be called if @bh is guaranteed to not be NULL.
1222 */
1223void __brelse(struct buffer_head *bh)
1224{
1225 if (atomic_read(&bh->b_count)) {
1226 put_bh(bh);
1227 return;
1228 }
1229 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1230}
1231EXPORT_SYMBOL(__brelse);
1232
1233/**
1234 * __bforget - Discard any dirty data in a buffer.
1235 * @bh: The buffer to forget.
1236 *
1237 * This variant of bforget() can be called if @bh is guaranteed to not
1238 * be NULL.
1239 */
1240void __bforget(struct buffer_head *bh)
1241{
1242 clear_buffer_dirty(bh);
1243 if (bh->b_assoc_map) {
1244 struct address_space *buffer_mapping = bh->b_folio->mapping;
1245
1246 spin_lock(&buffer_mapping->i_private_lock);
1247 list_del_init(&bh->b_assoc_buffers);
1248 bh->b_assoc_map = NULL;
1249 spin_unlock(&buffer_mapping->i_private_lock);
1250 }
1251 __brelse(bh);
1252}
1253EXPORT_SYMBOL(__bforget);
1254
1255static struct buffer_head *__bread_slow(struct buffer_head *bh)
1256{
1257 lock_buffer(bh);
1258 if (buffer_uptodate(bh)) {
1259 unlock_buffer(bh);
1260 return bh;
1261 } else {
1262 get_bh(bh);
1263 bh->b_end_io = end_buffer_read_sync;
1264 submit_bh(REQ_OP_READ, bh);
1265 wait_on_buffer(bh);
1266 if (buffer_uptodate(bh))
1267 return bh;
1268 }
1269 brelse(bh);
1270 return NULL;
1271}
1272
1273/*
1274 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1275 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1276 * refcount elevated by one when they're in an LRU. A buffer can only appear
1277 * once in a particular CPU's LRU. A single buffer can be present in multiple
1278 * CPU's LRUs at the same time.
1279 *
1280 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1281 * sb_find_get_block().
1282 *
1283 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1284 * a local interrupt disable for that.
1285 */
1286
1287#define BH_LRU_SIZE 16
1288
1289struct bh_lru {
1290 struct buffer_head *bhs[BH_LRU_SIZE];
1291};
1292
1293static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1294
1295#ifdef CONFIG_SMP
1296#define bh_lru_lock() local_irq_disable()
1297#define bh_lru_unlock() local_irq_enable()
1298#else
1299#define bh_lru_lock() preempt_disable()
1300#define bh_lru_unlock() preempt_enable()
1301#endif
1302
1303static inline void check_irqs_on(void)
1304{
1305#ifdef irqs_disabled
1306 BUG_ON(irqs_disabled());
1307#endif
1308}
1309
1310/*
1311 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1312 * inserted at the front, and the buffer_head at the back if any is evicted.
1313 * Or, if already in the LRU it is moved to the front.
1314 */
1315static void bh_lru_install(struct buffer_head *bh)
1316{
1317 struct buffer_head *evictee = bh;
1318 struct bh_lru *b;
1319 int i;
1320
1321 check_irqs_on();
1322 bh_lru_lock();
1323
1324 /*
1325 * the refcount of buffer_head in bh_lru prevents dropping the
1326 * attached page(i.e., try_to_free_buffers) so it could cause
1327 * failing page migration.
1328 * Skip putting upcoming bh into bh_lru until migration is done.
1329 */
1330 if (lru_cache_disabled() || cpu_is_isolated(smp_processor_id())) {
1331 bh_lru_unlock();
1332 return;
1333 }
1334
1335 b = this_cpu_ptr(&bh_lrus);
1336 for (i = 0; i < BH_LRU_SIZE; i++) {
1337 swap(evictee, b->bhs[i]);
1338 if (evictee == bh) {
1339 bh_lru_unlock();
1340 return;
1341 }
1342 }
1343
1344 get_bh(bh);
1345 bh_lru_unlock();
1346 brelse(evictee);
1347}
1348
1349/*
1350 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1351 */
1352static struct buffer_head *
1353lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1354{
1355 struct buffer_head *ret = NULL;
1356 unsigned int i;
1357
1358 check_irqs_on();
1359 bh_lru_lock();
1360 if (cpu_is_isolated(smp_processor_id())) {
1361 bh_lru_unlock();
1362 return NULL;
1363 }
1364 for (i = 0; i < BH_LRU_SIZE; i++) {
1365 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1366
1367 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1368 bh->b_size == size) {
1369 if (i) {
1370 while (i) {
1371 __this_cpu_write(bh_lrus.bhs[i],
1372 __this_cpu_read(bh_lrus.bhs[i - 1]));
1373 i--;
1374 }
1375 __this_cpu_write(bh_lrus.bhs[0], bh);
1376 }
1377 get_bh(bh);
1378 ret = bh;
1379 break;
1380 }
1381 }
1382 bh_lru_unlock();
1383 return ret;
1384}
1385
1386/*
1387 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1388 * it in the LRU and mark it as accessed. If it is not present then return
1389 * NULL
1390 */
1391struct buffer_head *
1392__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1393{
1394 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1395
1396 if (bh == NULL) {
1397 /* __find_get_block_slow will mark the page accessed */
1398 bh = __find_get_block_slow(bdev, block);
1399 if (bh)
1400 bh_lru_install(bh);
1401 } else
1402 touch_buffer(bh);
1403
1404 return bh;
1405}
1406EXPORT_SYMBOL(__find_get_block);
1407
1408/**
1409 * bdev_getblk - Get a buffer_head in a block device's buffer cache.
1410 * @bdev: The block device.
1411 * @block: The block number.
1412 * @size: The size of buffer_heads for this @bdev.
1413 * @gfp: The memory allocation flags to use.
1414 *
1415 * The returned buffer head has its reference count incremented, but is
1416 * not locked. The caller should call brelse() when it has finished
1417 * with the buffer. The buffer may not be uptodate. If needed, the
1418 * caller can bring it uptodate either by reading it or overwriting it.
1419 *
1420 * Return: The buffer head, or NULL if memory could not be allocated.
1421 */
1422struct buffer_head *bdev_getblk(struct block_device *bdev, sector_t block,
1423 unsigned size, gfp_t gfp)
1424{
1425 struct buffer_head *bh = __find_get_block(bdev, block, size);
1426
1427 might_alloc(gfp);
1428 if (bh)
1429 return bh;
1430
1431 return __getblk_slow(bdev, block, size, gfp);
1432}
1433EXPORT_SYMBOL(bdev_getblk);
1434
1435/*
1436 * Do async read-ahead on a buffer..
1437 */
1438void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1439{
1440 struct buffer_head *bh = bdev_getblk(bdev, block, size,
1441 GFP_NOWAIT | __GFP_MOVABLE);
1442
1443 if (likely(bh)) {
1444 bh_readahead(bh, REQ_RAHEAD);
1445 brelse(bh);
1446 }
1447}
1448EXPORT_SYMBOL(__breadahead);
1449
1450/**
1451 * __bread_gfp() - Read a block.
1452 * @bdev: The block device to read from.
1453 * @block: Block number in units of block size.
1454 * @size: The block size of this device in bytes.
1455 * @gfp: Not page allocation flags; see below.
1456 *
1457 * You are not expected to call this function. You should use one of
1458 * sb_bread(), sb_bread_unmovable() or __bread().
1459 *
1460 * Read a specified block, and return the buffer head that refers to it.
1461 * If @gfp is 0, the memory will be allocated using the block device's
1462 * default GFP flags. If @gfp is __GFP_MOVABLE, the memory may be
1463 * allocated from a movable area. Do not pass in a complete set of
1464 * GFP flags.
1465 *
1466 * The returned buffer head has its refcount increased. The caller should
1467 * call brelse() when it has finished with the buffer.
1468 *
1469 * Context: May sleep waiting for I/O.
1470 * Return: NULL if the block was unreadable.
1471 */
1472struct buffer_head *__bread_gfp(struct block_device *bdev, sector_t block,
1473 unsigned size, gfp_t gfp)
1474{
1475 struct buffer_head *bh;
1476
1477 gfp |= mapping_gfp_constraint(bdev->bd_mapping, ~__GFP_FS);
1478
1479 /*
1480 * Prefer looping in the allocator rather than here, at least that
1481 * code knows what it's doing.
1482 */
1483 gfp |= __GFP_NOFAIL;
1484
1485 bh = bdev_getblk(bdev, block, size, gfp);
1486
1487 if (likely(bh) && !buffer_uptodate(bh))
1488 bh = __bread_slow(bh);
1489 return bh;
1490}
1491EXPORT_SYMBOL(__bread_gfp);
1492
1493static void __invalidate_bh_lrus(struct bh_lru *b)
1494{
1495 int i;
1496
1497 for (i = 0; i < BH_LRU_SIZE; i++) {
1498 brelse(b->bhs[i]);
1499 b->bhs[i] = NULL;
1500 }
1501}
1502/*
1503 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1504 * This doesn't race because it runs in each cpu either in irq
1505 * or with preempt disabled.
1506 */
1507static void invalidate_bh_lru(void *arg)
1508{
1509 struct bh_lru *b = &get_cpu_var(bh_lrus);
1510
1511 __invalidate_bh_lrus(b);
1512 put_cpu_var(bh_lrus);
1513}
1514
1515bool has_bh_in_lru(int cpu, void *dummy)
1516{
1517 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1518 int i;
1519
1520 for (i = 0; i < BH_LRU_SIZE; i++) {
1521 if (b->bhs[i])
1522 return true;
1523 }
1524
1525 return false;
1526}
1527
1528void invalidate_bh_lrus(void)
1529{
1530 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1531}
1532EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1533
1534/*
1535 * It's called from workqueue context so we need a bh_lru_lock to close
1536 * the race with preemption/irq.
1537 */
1538void invalidate_bh_lrus_cpu(void)
1539{
1540 struct bh_lru *b;
1541
1542 bh_lru_lock();
1543 b = this_cpu_ptr(&bh_lrus);
1544 __invalidate_bh_lrus(b);
1545 bh_lru_unlock();
1546}
1547
1548void folio_set_bh(struct buffer_head *bh, struct folio *folio,
1549 unsigned long offset)
1550{
1551 bh->b_folio = folio;
1552 BUG_ON(offset >= folio_size(folio));
1553 if (folio_test_highmem(folio))
1554 /*
1555 * This catches illegal uses and preserves the offset:
1556 */
1557 bh->b_data = (char *)(0 + offset);
1558 else
1559 bh->b_data = folio_address(folio) + offset;
1560}
1561EXPORT_SYMBOL(folio_set_bh);
1562
1563/*
1564 * Called when truncating a buffer on a page completely.
1565 */
1566
1567/* Bits that are cleared during an invalidate */
1568#define BUFFER_FLAGS_DISCARD \
1569 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1570 1 << BH_Delay | 1 << BH_Unwritten)
1571
1572static void discard_buffer(struct buffer_head * bh)
1573{
1574 unsigned long b_state;
1575
1576 lock_buffer(bh);
1577 clear_buffer_dirty(bh);
1578 bh->b_bdev = NULL;
1579 b_state = READ_ONCE(bh->b_state);
1580 do {
1581 } while (!try_cmpxchg(&bh->b_state, &b_state,
1582 b_state & ~BUFFER_FLAGS_DISCARD));
1583 unlock_buffer(bh);
1584}
1585
1586/**
1587 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1588 * @folio: The folio which is affected.
1589 * @offset: start of the range to invalidate
1590 * @length: length of the range to invalidate
1591 *
1592 * block_invalidate_folio() is called when all or part of the folio has been
1593 * invalidated by a truncate operation.
1594 *
1595 * block_invalidate_folio() does not have to release all buffers, but it must
1596 * ensure that no dirty buffer is left outside @offset and that no I/O
1597 * is underway against any of the blocks which are outside the truncation
1598 * point. Because the caller is about to free (and possibly reuse) those
1599 * blocks on-disk.
1600 */
1601void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1602{
1603 struct buffer_head *head, *bh, *next;
1604 size_t curr_off = 0;
1605 size_t stop = length + offset;
1606
1607 BUG_ON(!folio_test_locked(folio));
1608
1609 /*
1610 * Check for overflow
1611 */
1612 BUG_ON(stop > folio_size(folio) || stop < length);
1613
1614 head = folio_buffers(folio);
1615 if (!head)
1616 return;
1617
1618 bh = head;
1619 do {
1620 size_t next_off = curr_off + bh->b_size;
1621 next = bh->b_this_page;
1622
1623 /*
1624 * Are we still fully in range ?
1625 */
1626 if (next_off > stop)
1627 goto out;
1628
1629 /*
1630 * is this block fully invalidated?
1631 */
1632 if (offset <= curr_off)
1633 discard_buffer(bh);
1634 curr_off = next_off;
1635 bh = next;
1636 } while (bh != head);
1637
1638 /*
1639 * We release buffers only if the entire folio is being invalidated.
1640 * The get_block cached value has been unconditionally invalidated,
1641 * so real IO is not possible anymore.
1642 */
1643 if (length == folio_size(folio))
1644 filemap_release_folio(folio, 0);
1645out:
1646 folio_clear_mappedtodisk(folio);
1647 return;
1648}
1649EXPORT_SYMBOL(block_invalidate_folio);
1650
1651/*
1652 * We attach and possibly dirty the buffers atomically wrt
1653 * block_dirty_folio() via i_private_lock. try_to_free_buffers
1654 * is already excluded via the folio lock.
1655 */
1656struct buffer_head *create_empty_buffers(struct folio *folio,
1657 unsigned long blocksize, unsigned long b_state)
1658{
1659 struct buffer_head *bh, *head, *tail;
1660 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT | __GFP_NOFAIL;
1661
1662 head = folio_alloc_buffers(folio, blocksize, gfp);
1663 bh = head;
1664 do {
1665 bh->b_state |= b_state;
1666 tail = bh;
1667 bh = bh->b_this_page;
1668 } while (bh);
1669 tail->b_this_page = head;
1670
1671 spin_lock(&folio->mapping->i_private_lock);
1672 if (folio_test_uptodate(folio) || folio_test_dirty(folio)) {
1673 bh = head;
1674 do {
1675 if (folio_test_dirty(folio))
1676 set_buffer_dirty(bh);
1677 if (folio_test_uptodate(folio))
1678 set_buffer_uptodate(bh);
1679 bh = bh->b_this_page;
1680 } while (bh != head);
1681 }
1682 folio_attach_private(folio, head);
1683 spin_unlock(&folio->mapping->i_private_lock);
1684
1685 return head;
1686}
1687EXPORT_SYMBOL(create_empty_buffers);
1688
1689/**
1690 * clean_bdev_aliases: clean a range of buffers in block device
1691 * @bdev: Block device to clean buffers in
1692 * @block: Start of a range of blocks to clean
1693 * @len: Number of blocks to clean
1694 *
1695 * We are taking a range of blocks for data and we don't want writeback of any
1696 * buffer-cache aliases starting from return from this function and until the
1697 * moment when something will explicitly mark the buffer dirty (hopefully that
1698 * will not happen until we will free that block ;-) We don't even need to mark
1699 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1700 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1701 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1702 * would confuse anyone who might pick it with bread() afterwards...
1703 *
1704 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1705 * writeout I/O going on against recently-freed buffers. We don't wait on that
1706 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1707 * need to. That happens here.
1708 */
1709void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1710{
1711 struct address_space *bd_mapping = bdev->bd_mapping;
1712 const int blkbits = bd_mapping->host->i_blkbits;
1713 struct folio_batch fbatch;
1714 pgoff_t index = ((loff_t)block << blkbits) / PAGE_SIZE;
1715 pgoff_t end;
1716 int i, count;
1717 struct buffer_head *bh;
1718 struct buffer_head *head;
1719
1720 end = ((loff_t)(block + len - 1) << blkbits) / PAGE_SIZE;
1721 folio_batch_init(&fbatch);
1722 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1723 count = folio_batch_count(&fbatch);
1724 for (i = 0; i < count; i++) {
1725 struct folio *folio = fbatch.folios[i];
1726
1727 if (!folio_buffers(folio))
1728 continue;
1729 /*
1730 * We use folio lock instead of bd_mapping->i_private_lock
1731 * to pin buffers here since we can afford to sleep and
1732 * it scales better than a global spinlock lock.
1733 */
1734 folio_lock(folio);
1735 /* Recheck when the folio is locked which pins bhs */
1736 head = folio_buffers(folio);
1737 if (!head)
1738 goto unlock_page;
1739 bh = head;
1740 do {
1741 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1742 goto next;
1743 if (bh->b_blocknr >= block + len)
1744 break;
1745 clear_buffer_dirty(bh);
1746 wait_on_buffer(bh);
1747 clear_buffer_req(bh);
1748next:
1749 bh = bh->b_this_page;
1750 } while (bh != head);
1751unlock_page:
1752 folio_unlock(folio);
1753 }
1754 folio_batch_release(&fbatch);
1755 cond_resched();
1756 /* End of range already reached? */
1757 if (index > end || !index)
1758 break;
1759 }
1760}
1761EXPORT_SYMBOL(clean_bdev_aliases);
1762
1763static struct buffer_head *folio_create_buffers(struct folio *folio,
1764 struct inode *inode,
1765 unsigned int b_state)
1766{
1767 struct buffer_head *bh;
1768
1769 BUG_ON(!folio_test_locked(folio));
1770
1771 bh = folio_buffers(folio);
1772 if (!bh)
1773 bh = create_empty_buffers(folio,
1774 1 << READ_ONCE(inode->i_blkbits), b_state);
1775 return bh;
1776}
1777
1778/*
1779 * NOTE! All mapped/uptodate combinations are valid:
1780 *
1781 * Mapped Uptodate Meaning
1782 *
1783 * No No "unknown" - must do get_block()
1784 * No Yes "hole" - zero-filled
1785 * Yes No "allocated" - allocated on disk, not read in
1786 * Yes Yes "valid" - allocated and up-to-date in memory.
1787 *
1788 * "Dirty" is valid only with the last case (mapped+uptodate).
1789 */
1790
1791/*
1792 * While block_write_full_folio is writing back the dirty buffers under
1793 * the page lock, whoever dirtied the buffers may decide to clean them
1794 * again at any time. We handle that by only looking at the buffer
1795 * state inside lock_buffer().
1796 *
1797 * If block_write_full_folio() is called for regular writeback
1798 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1799 * locked buffer. This only can happen if someone has written the buffer
1800 * directly, with submit_bh(). At the address_space level PageWriteback
1801 * prevents this contention from occurring.
1802 *
1803 * If block_write_full_folio() is called with wbc->sync_mode ==
1804 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1805 * causes the writes to be flagged as synchronous writes.
1806 */
1807int __block_write_full_folio(struct inode *inode, struct folio *folio,
1808 get_block_t *get_block, struct writeback_control *wbc)
1809{
1810 int err;
1811 sector_t block;
1812 sector_t last_block;
1813 struct buffer_head *bh, *head;
1814 size_t blocksize;
1815 int nr_underway = 0;
1816 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1817
1818 head = folio_create_buffers(folio, inode,
1819 (1 << BH_Dirty) | (1 << BH_Uptodate));
1820
1821 /*
1822 * Be very careful. We have no exclusion from block_dirty_folio
1823 * here, and the (potentially unmapped) buffers may become dirty at
1824 * any time. If a buffer becomes dirty here after we've inspected it
1825 * then we just miss that fact, and the folio stays dirty.
1826 *
1827 * Buffers outside i_size may be dirtied by block_dirty_folio;
1828 * handle that here by just cleaning them.
1829 */
1830
1831 bh = head;
1832 blocksize = bh->b_size;
1833
1834 block = div_u64(folio_pos(folio), blocksize);
1835 last_block = div_u64(i_size_read(inode) - 1, blocksize);
1836
1837 /*
1838 * Get all the dirty buffers mapped to disk addresses and
1839 * handle any aliases from the underlying blockdev's mapping.
1840 */
1841 do {
1842 if (block > last_block) {
1843 /*
1844 * mapped buffers outside i_size will occur, because
1845 * this folio can be outside i_size when there is a
1846 * truncate in progress.
1847 */
1848 /*
1849 * The buffer was zeroed by block_write_full_folio()
1850 */
1851 clear_buffer_dirty(bh);
1852 set_buffer_uptodate(bh);
1853 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1854 buffer_dirty(bh)) {
1855 WARN_ON(bh->b_size != blocksize);
1856 err = get_block(inode, block, bh, 1);
1857 if (err)
1858 goto recover;
1859 clear_buffer_delay(bh);
1860 if (buffer_new(bh)) {
1861 /* blockdev mappings never come here */
1862 clear_buffer_new(bh);
1863 clean_bdev_bh_alias(bh);
1864 }
1865 }
1866 bh = bh->b_this_page;
1867 block++;
1868 } while (bh != head);
1869
1870 do {
1871 if (!buffer_mapped(bh))
1872 continue;
1873 /*
1874 * If it's a fully non-blocking write attempt and we cannot
1875 * lock the buffer then redirty the folio. Note that this can
1876 * potentially cause a busy-wait loop from writeback threads
1877 * and kswapd activity, but those code paths have their own
1878 * higher-level throttling.
1879 */
1880 if (wbc->sync_mode != WB_SYNC_NONE) {
1881 lock_buffer(bh);
1882 } else if (!trylock_buffer(bh)) {
1883 folio_redirty_for_writepage(wbc, folio);
1884 continue;
1885 }
1886 if (test_clear_buffer_dirty(bh)) {
1887 mark_buffer_async_write_endio(bh,
1888 end_buffer_async_write);
1889 } else {
1890 unlock_buffer(bh);
1891 }
1892 } while ((bh = bh->b_this_page) != head);
1893
1894 /*
1895 * The folio and its buffers are protected by the writeback flag,
1896 * so we can drop the bh refcounts early.
1897 */
1898 BUG_ON(folio_test_writeback(folio));
1899 folio_start_writeback(folio);
1900
1901 do {
1902 struct buffer_head *next = bh->b_this_page;
1903 if (buffer_async_write(bh)) {
1904 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1905 inode->i_write_hint, wbc);
1906 nr_underway++;
1907 }
1908 bh = next;
1909 } while (bh != head);
1910 folio_unlock(folio);
1911
1912 err = 0;
1913done:
1914 if (nr_underway == 0) {
1915 /*
1916 * The folio was marked dirty, but the buffers were
1917 * clean. Someone wrote them back by hand with
1918 * write_dirty_buffer/submit_bh. A rare case.
1919 */
1920 folio_end_writeback(folio);
1921
1922 /*
1923 * The folio and buffer_heads can be released at any time from
1924 * here on.
1925 */
1926 }
1927 return err;
1928
1929recover:
1930 /*
1931 * ENOSPC, or some other error. We may already have added some
1932 * blocks to the file, so we need to write these out to avoid
1933 * exposing stale data.
1934 * The folio is currently locked and not marked for writeback
1935 */
1936 bh = head;
1937 /* Recovery: lock and submit the mapped buffers */
1938 do {
1939 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1940 !buffer_delay(bh)) {
1941 lock_buffer(bh);
1942 mark_buffer_async_write_endio(bh,
1943 end_buffer_async_write);
1944 } else {
1945 /*
1946 * The buffer may have been set dirty during
1947 * attachment to a dirty folio.
1948 */
1949 clear_buffer_dirty(bh);
1950 }
1951 } while ((bh = bh->b_this_page) != head);
1952 BUG_ON(folio_test_writeback(folio));
1953 mapping_set_error(folio->mapping, err);
1954 folio_start_writeback(folio);
1955 do {
1956 struct buffer_head *next = bh->b_this_page;
1957 if (buffer_async_write(bh)) {
1958 clear_buffer_dirty(bh);
1959 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1960 inode->i_write_hint, wbc);
1961 nr_underway++;
1962 }
1963 bh = next;
1964 } while (bh != head);
1965 folio_unlock(folio);
1966 goto done;
1967}
1968EXPORT_SYMBOL(__block_write_full_folio);
1969
1970/*
1971 * If a folio has any new buffers, zero them out here, and mark them uptodate
1972 * and dirty so they'll be written out (in order to prevent uninitialised
1973 * block data from leaking). And clear the new bit.
1974 */
1975void folio_zero_new_buffers(struct folio *folio, size_t from, size_t to)
1976{
1977 size_t block_start, block_end;
1978 struct buffer_head *head, *bh;
1979
1980 BUG_ON(!folio_test_locked(folio));
1981 head = folio_buffers(folio);
1982 if (!head)
1983 return;
1984
1985 bh = head;
1986 block_start = 0;
1987 do {
1988 block_end = block_start + bh->b_size;
1989
1990 if (buffer_new(bh)) {
1991 if (block_end > from && block_start < to) {
1992 if (!folio_test_uptodate(folio)) {
1993 size_t start, xend;
1994
1995 start = max(from, block_start);
1996 xend = min(to, block_end);
1997
1998 folio_zero_segment(folio, start, xend);
1999 set_buffer_uptodate(bh);
2000 }
2001
2002 clear_buffer_new(bh);
2003 mark_buffer_dirty(bh);
2004 }
2005 }
2006
2007 block_start = block_end;
2008 bh = bh->b_this_page;
2009 } while (bh != head);
2010}
2011EXPORT_SYMBOL(folio_zero_new_buffers);
2012
2013static int
2014iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
2015 const struct iomap *iomap)
2016{
2017 loff_t offset = (loff_t)block << inode->i_blkbits;
2018
2019 bh->b_bdev = iomap->bdev;
2020
2021 /*
2022 * Block points to offset in file we need to map, iomap contains
2023 * the offset at which the map starts. If the map ends before the
2024 * current block, then do not map the buffer and let the caller
2025 * handle it.
2026 */
2027 if (offset >= iomap->offset + iomap->length)
2028 return -EIO;
2029
2030 switch (iomap->type) {
2031 case IOMAP_HOLE:
2032 /*
2033 * If the buffer is not up to date or beyond the current EOF,
2034 * we need to mark it as new to ensure sub-block zeroing is
2035 * executed if necessary.
2036 */
2037 if (!buffer_uptodate(bh) ||
2038 (offset >= i_size_read(inode)))
2039 set_buffer_new(bh);
2040 return 0;
2041 case IOMAP_DELALLOC:
2042 if (!buffer_uptodate(bh) ||
2043 (offset >= i_size_read(inode)))
2044 set_buffer_new(bh);
2045 set_buffer_uptodate(bh);
2046 set_buffer_mapped(bh);
2047 set_buffer_delay(bh);
2048 return 0;
2049 case IOMAP_UNWRITTEN:
2050 /*
2051 * For unwritten regions, we always need to ensure that regions
2052 * in the block we are not writing to are zeroed. Mark the
2053 * buffer as new to ensure this.
2054 */
2055 set_buffer_new(bh);
2056 set_buffer_unwritten(bh);
2057 fallthrough;
2058 case IOMAP_MAPPED:
2059 if ((iomap->flags & IOMAP_F_NEW) ||
2060 offset >= i_size_read(inode)) {
2061 /*
2062 * This can happen if truncating the block device races
2063 * with the check in the caller as i_size updates on
2064 * block devices aren't synchronized by i_rwsem for
2065 * block devices.
2066 */
2067 if (S_ISBLK(inode->i_mode))
2068 return -EIO;
2069 set_buffer_new(bh);
2070 }
2071 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
2072 inode->i_blkbits;
2073 set_buffer_mapped(bh);
2074 return 0;
2075 default:
2076 WARN_ON_ONCE(1);
2077 return -EIO;
2078 }
2079}
2080
2081int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
2082 get_block_t *get_block, const struct iomap *iomap)
2083{
2084 size_t from = offset_in_folio(folio, pos);
2085 size_t to = from + len;
2086 struct inode *inode = folio->mapping->host;
2087 size_t block_start, block_end;
2088 sector_t block;
2089 int err = 0;
2090 size_t blocksize;
2091 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2092
2093 BUG_ON(!folio_test_locked(folio));
2094 BUG_ON(to > folio_size(folio));
2095 BUG_ON(from > to);
2096
2097 head = folio_create_buffers(folio, inode, 0);
2098 blocksize = head->b_size;
2099 block = div_u64(folio_pos(folio), blocksize);
2100
2101 for (bh = head, block_start = 0; bh != head || !block_start;
2102 block++, block_start=block_end, bh = bh->b_this_page) {
2103 block_end = block_start + blocksize;
2104 if (block_end <= from || block_start >= to) {
2105 if (folio_test_uptodate(folio)) {
2106 if (!buffer_uptodate(bh))
2107 set_buffer_uptodate(bh);
2108 }
2109 continue;
2110 }
2111 if (buffer_new(bh))
2112 clear_buffer_new(bh);
2113 if (!buffer_mapped(bh)) {
2114 WARN_ON(bh->b_size != blocksize);
2115 if (get_block)
2116 err = get_block(inode, block, bh, 1);
2117 else
2118 err = iomap_to_bh(inode, block, bh, iomap);
2119 if (err)
2120 break;
2121
2122 if (buffer_new(bh)) {
2123 clean_bdev_bh_alias(bh);
2124 if (folio_test_uptodate(folio)) {
2125 clear_buffer_new(bh);
2126 set_buffer_uptodate(bh);
2127 mark_buffer_dirty(bh);
2128 continue;
2129 }
2130 if (block_end > to || block_start < from)
2131 folio_zero_segments(folio,
2132 to, block_end,
2133 block_start, from);
2134 continue;
2135 }
2136 }
2137 if (folio_test_uptodate(folio)) {
2138 if (!buffer_uptodate(bh))
2139 set_buffer_uptodate(bh);
2140 continue;
2141 }
2142 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2143 !buffer_unwritten(bh) &&
2144 (block_start < from || block_end > to)) {
2145 bh_read_nowait(bh, 0);
2146 *wait_bh++=bh;
2147 }
2148 }
2149 /*
2150 * If we issued read requests - let them complete.
2151 */
2152 while(wait_bh > wait) {
2153 wait_on_buffer(*--wait_bh);
2154 if (!buffer_uptodate(*wait_bh))
2155 err = -EIO;
2156 }
2157 if (unlikely(err))
2158 folio_zero_new_buffers(folio, from, to);
2159 return err;
2160}
2161
2162int __block_write_begin(struct folio *folio, loff_t pos, unsigned len,
2163 get_block_t *get_block)
2164{
2165 return __block_write_begin_int(folio, pos, len, get_block, NULL);
2166}
2167EXPORT_SYMBOL(__block_write_begin);
2168
2169static void __block_commit_write(struct folio *folio, size_t from, size_t to)
2170{
2171 size_t block_start, block_end;
2172 bool partial = false;
2173 unsigned blocksize;
2174 struct buffer_head *bh, *head;
2175
2176 bh = head = folio_buffers(folio);
2177 if (!bh)
2178 return;
2179 blocksize = bh->b_size;
2180
2181 block_start = 0;
2182 do {
2183 block_end = block_start + blocksize;
2184 if (block_end <= from || block_start >= to) {
2185 if (!buffer_uptodate(bh))
2186 partial = true;
2187 } else {
2188 set_buffer_uptodate(bh);
2189 mark_buffer_dirty(bh);
2190 }
2191 if (buffer_new(bh))
2192 clear_buffer_new(bh);
2193
2194 block_start = block_end;
2195 bh = bh->b_this_page;
2196 } while (bh != head);
2197
2198 /*
2199 * If this is a partial write which happened to make all buffers
2200 * uptodate then we can optimize away a bogus read_folio() for
2201 * the next read(). Here we 'discover' whether the folio went
2202 * uptodate as a result of this (potentially partial) write.
2203 */
2204 if (!partial)
2205 folio_mark_uptodate(folio);
2206}
2207
2208/*
2209 * block_write_begin takes care of the basic task of block allocation and
2210 * bringing partial write blocks uptodate first.
2211 *
2212 * The filesystem needs to handle block truncation upon failure.
2213 */
2214int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2215 struct folio **foliop, get_block_t *get_block)
2216{
2217 pgoff_t index = pos >> PAGE_SHIFT;
2218 struct folio *folio;
2219 int status;
2220
2221 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
2222 mapping_gfp_mask(mapping));
2223 if (IS_ERR(folio))
2224 return PTR_ERR(folio);
2225
2226 status = __block_write_begin_int(folio, pos, len, get_block, NULL);
2227 if (unlikely(status)) {
2228 folio_unlock(folio);
2229 folio_put(folio);
2230 folio = NULL;
2231 }
2232
2233 *foliop = folio;
2234 return status;
2235}
2236EXPORT_SYMBOL(block_write_begin);
2237
2238int block_write_end(struct file *file, struct address_space *mapping,
2239 loff_t pos, unsigned len, unsigned copied,
2240 struct folio *folio, void *fsdata)
2241{
2242 size_t start = pos - folio_pos(folio);
2243
2244 if (unlikely(copied < len)) {
2245 /*
2246 * The buffers that were written will now be uptodate, so
2247 * we don't have to worry about a read_folio reading them
2248 * and overwriting a partial write. However if we have
2249 * encountered a short write and only partially written
2250 * into a buffer, it will not be marked uptodate, so a
2251 * read_folio might come in and destroy our partial write.
2252 *
2253 * Do the simplest thing, and just treat any short write to a
2254 * non uptodate folio as a zero-length write, and force the
2255 * caller to redo the whole thing.
2256 */
2257 if (!folio_test_uptodate(folio))
2258 copied = 0;
2259
2260 folio_zero_new_buffers(folio, start+copied, start+len);
2261 }
2262 flush_dcache_folio(folio);
2263
2264 /* This could be a short (even 0-length) commit */
2265 __block_commit_write(folio, start, start + copied);
2266
2267 return copied;
2268}
2269EXPORT_SYMBOL(block_write_end);
2270
2271int generic_write_end(struct file *file, struct address_space *mapping,
2272 loff_t pos, unsigned len, unsigned copied,
2273 struct folio *folio, void *fsdata)
2274{
2275 struct inode *inode = mapping->host;
2276 loff_t old_size = inode->i_size;
2277 bool i_size_changed = false;
2278
2279 copied = block_write_end(file, mapping, pos, len, copied, folio, fsdata);
2280
2281 /*
2282 * No need to use i_size_read() here, the i_size cannot change under us
2283 * because we hold i_rwsem.
2284 *
2285 * But it's important to update i_size while still holding folio lock:
2286 * page writeout could otherwise come in and zero beyond i_size.
2287 */
2288 if (pos + copied > inode->i_size) {
2289 i_size_write(inode, pos + copied);
2290 i_size_changed = true;
2291 }
2292
2293 folio_unlock(folio);
2294 folio_put(folio);
2295
2296 if (old_size < pos)
2297 pagecache_isize_extended(inode, old_size, pos);
2298 /*
2299 * Don't mark the inode dirty under page lock. First, it unnecessarily
2300 * makes the holding time of page lock longer. Second, it forces lock
2301 * ordering of page lock and transaction start for journaling
2302 * filesystems.
2303 */
2304 if (i_size_changed)
2305 mark_inode_dirty(inode);
2306 return copied;
2307}
2308EXPORT_SYMBOL(generic_write_end);
2309
2310/*
2311 * block_is_partially_uptodate checks whether buffers within a folio are
2312 * uptodate or not.
2313 *
2314 * Returns true if all buffers which correspond to the specified part
2315 * of the folio are uptodate.
2316 */
2317bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2318{
2319 unsigned block_start, block_end, blocksize;
2320 unsigned to;
2321 struct buffer_head *bh, *head;
2322 bool ret = true;
2323
2324 head = folio_buffers(folio);
2325 if (!head)
2326 return false;
2327 blocksize = head->b_size;
2328 to = min_t(unsigned, folio_size(folio) - from, count);
2329 to = from + to;
2330 if (from < blocksize && to > folio_size(folio) - blocksize)
2331 return false;
2332
2333 bh = head;
2334 block_start = 0;
2335 do {
2336 block_end = block_start + blocksize;
2337 if (block_end > from && block_start < to) {
2338 if (!buffer_uptodate(bh)) {
2339 ret = false;
2340 break;
2341 }
2342 if (block_end >= to)
2343 break;
2344 }
2345 block_start = block_end;
2346 bh = bh->b_this_page;
2347 } while (bh != head);
2348
2349 return ret;
2350}
2351EXPORT_SYMBOL(block_is_partially_uptodate);
2352
2353/*
2354 * Generic "read_folio" function for block devices that have the normal
2355 * get_block functionality. This is most of the block device filesystems.
2356 * Reads the folio asynchronously --- the unlock_buffer() and
2357 * set/clear_buffer_uptodate() functions propagate buffer state into the
2358 * folio once IO has completed.
2359 */
2360int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2361{
2362 struct inode *inode = folio->mapping->host;
2363 sector_t iblock, lblock;
2364 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2365 size_t blocksize;
2366 int nr, i;
2367 int fully_mapped = 1;
2368 bool page_error = false;
2369 loff_t limit = i_size_read(inode);
2370
2371 /* This is needed for ext4. */
2372 if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
2373 limit = inode->i_sb->s_maxbytes;
2374
2375 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2376
2377 head = folio_create_buffers(folio, inode, 0);
2378 blocksize = head->b_size;
2379
2380 iblock = div_u64(folio_pos(folio), blocksize);
2381 lblock = div_u64(limit + blocksize - 1, blocksize);
2382 bh = head;
2383 nr = 0;
2384 i = 0;
2385
2386 do {
2387 if (buffer_uptodate(bh))
2388 continue;
2389
2390 if (!buffer_mapped(bh)) {
2391 int err = 0;
2392
2393 fully_mapped = 0;
2394 if (iblock < lblock) {
2395 WARN_ON(bh->b_size != blocksize);
2396 err = get_block(inode, iblock, bh, 0);
2397 if (err)
2398 page_error = true;
2399 }
2400 if (!buffer_mapped(bh)) {
2401 folio_zero_range(folio, i * blocksize,
2402 blocksize);
2403 if (!err)
2404 set_buffer_uptodate(bh);
2405 continue;
2406 }
2407 /*
2408 * get_block() might have updated the buffer
2409 * synchronously
2410 */
2411 if (buffer_uptodate(bh))
2412 continue;
2413 }
2414 arr[nr++] = bh;
2415 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2416
2417 if (fully_mapped)
2418 folio_set_mappedtodisk(folio);
2419
2420 if (!nr) {
2421 /*
2422 * All buffers are uptodate or get_block() returned an
2423 * error when trying to map them - we can finish the read.
2424 */
2425 folio_end_read(folio, !page_error);
2426 return 0;
2427 }
2428
2429 /* Stage two: lock the buffers */
2430 for (i = 0; i < nr; i++) {
2431 bh = arr[i];
2432 lock_buffer(bh);
2433 mark_buffer_async_read(bh);
2434 }
2435
2436 /*
2437 * Stage 3: start the IO. Check for uptodateness
2438 * inside the buffer lock in case another process reading
2439 * the underlying blockdev brought it uptodate (the sct fix).
2440 */
2441 for (i = 0; i < nr; i++) {
2442 bh = arr[i];
2443 if (buffer_uptodate(bh))
2444 end_buffer_async_read(bh, 1);
2445 else
2446 submit_bh(REQ_OP_READ, bh);
2447 }
2448 return 0;
2449}
2450EXPORT_SYMBOL(block_read_full_folio);
2451
2452/* utility function for filesystems that need to do work on expanding
2453 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2454 * deal with the hole.
2455 */
2456int generic_cont_expand_simple(struct inode *inode, loff_t size)
2457{
2458 struct address_space *mapping = inode->i_mapping;
2459 const struct address_space_operations *aops = mapping->a_ops;
2460 struct folio *folio;
2461 void *fsdata = NULL;
2462 int err;
2463
2464 err = inode_newsize_ok(inode, size);
2465 if (err)
2466 goto out;
2467
2468 err = aops->write_begin(NULL, mapping, size, 0, &folio, &fsdata);
2469 if (err)
2470 goto out;
2471
2472 err = aops->write_end(NULL, mapping, size, 0, 0, folio, fsdata);
2473 BUG_ON(err > 0);
2474
2475out:
2476 return err;
2477}
2478EXPORT_SYMBOL(generic_cont_expand_simple);
2479
2480static int cont_expand_zero(struct file *file, struct address_space *mapping,
2481 loff_t pos, loff_t *bytes)
2482{
2483 struct inode *inode = mapping->host;
2484 const struct address_space_operations *aops = mapping->a_ops;
2485 unsigned int blocksize = i_blocksize(inode);
2486 struct folio *folio;
2487 void *fsdata = NULL;
2488 pgoff_t index, curidx;
2489 loff_t curpos;
2490 unsigned zerofrom, offset, len;
2491 int err = 0;
2492
2493 index = pos >> PAGE_SHIFT;
2494 offset = pos & ~PAGE_MASK;
2495
2496 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2497 zerofrom = curpos & ~PAGE_MASK;
2498 if (zerofrom & (blocksize-1)) {
2499 *bytes |= (blocksize-1);
2500 (*bytes)++;
2501 }
2502 len = PAGE_SIZE - zerofrom;
2503
2504 err = aops->write_begin(file, mapping, curpos, len,
2505 &folio, &fsdata);
2506 if (err)
2507 goto out;
2508 folio_zero_range(folio, offset_in_folio(folio, curpos), len);
2509 err = aops->write_end(file, mapping, curpos, len, len,
2510 folio, fsdata);
2511 if (err < 0)
2512 goto out;
2513 BUG_ON(err != len);
2514 err = 0;
2515
2516 balance_dirty_pages_ratelimited(mapping);
2517
2518 if (fatal_signal_pending(current)) {
2519 err = -EINTR;
2520 goto out;
2521 }
2522 }
2523
2524 /* page covers the boundary, find the boundary offset */
2525 if (index == curidx) {
2526 zerofrom = curpos & ~PAGE_MASK;
2527 /* if we will expand the thing last block will be filled */
2528 if (offset <= zerofrom) {
2529 goto out;
2530 }
2531 if (zerofrom & (blocksize-1)) {
2532 *bytes |= (blocksize-1);
2533 (*bytes)++;
2534 }
2535 len = offset - zerofrom;
2536
2537 err = aops->write_begin(file, mapping, curpos, len,
2538 &folio, &fsdata);
2539 if (err)
2540 goto out;
2541 folio_zero_range(folio, offset_in_folio(folio, curpos), len);
2542 err = aops->write_end(file, mapping, curpos, len, len,
2543 folio, fsdata);
2544 if (err < 0)
2545 goto out;
2546 BUG_ON(err != len);
2547 err = 0;
2548 }
2549out:
2550 return err;
2551}
2552
2553/*
2554 * For moronic filesystems that do not allow holes in file.
2555 * We may have to extend the file.
2556 */
2557int cont_write_begin(struct file *file, struct address_space *mapping,
2558 loff_t pos, unsigned len,
2559 struct folio **foliop, void **fsdata,
2560 get_block_t *get_block, loff_t *bytes)
2561{
2562 struct inode *inode = mapping->host;
2563 unsigned int blocksize = i_blocksize(inode);
2564 unsigned int zerofrom;
2565 int err;
2566
2567 err = cont_expand_zero(file, mapping, pos, bytes);
2568 if (err)
2569 return err;
2570
2571 zerofrom = *bytes & ~PAGE_MASK;
2572 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2573 *bytes |= (blocksize-1);
2574 (*bytes)++;
2575 }
2576
2577 return block_write_begin(mapping, pos, len, foliop, get_block);
2578}
2579EXPORT_SYMBOL(cont_write_begin);
2580
2581void block_commit_write(struct page *page, unsigned from, unsigned to)
2582{
2583 struct folio *folio = page_folio(page);
2584 __block_commit_write(folio, from, to);
2585}
2586EXPORT_SYMBOL(block_commit_write);
2587
2588/*
2589 * block_page_mkwrite() is not allowed to change the file size as it gets
2590 * called from a page fault handler when a page is first dirtied. Hence we must
2591 * be careful to check for EOF conditions here. We set the page up correctly
2592 * for a written page which means we get ENOSPC checking when writing into
2593 * holes and correct delalloc and unwritten extent mapping on filesystems that
2594 * support these features.
2595 *
2596 * We are not allowed to take the i_mutex here so we have to play games to
2597 * protect against truncate races as the page could now be beyond EOF. Because
2598 * truncate writes the inode size before removing pages, once we have the
2599 * page lock we can determine safely if the page is beyond EOF. If it is not
2600 * beyond EOF, then the page is guaranteed safe against truncation until we
2601 * unlock the page.
2602 *
2603 * Direct callers of this function should protect against filesystem freezing
2604 * using sb_start_pagefault() - sb_end_pagefault() functions.
2605 */
2606int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2607 get_block_t get_block)
2608{
2609 struct folio *folio = page_folio(vmf->page);
2610 struct inode *inode = file_inode(vma->vm_file);
2611 unsigned long end;
2612 loff_t size;
2613 int ret;
2614
2615 folio_lock(folio);
2616 size = i_size_read(inode);
2617 if ((folio->mapping != inode->i_mapping) ||
2618 (folio_pos(folio) >= size)) {
2619 /* We overload EFAULT to mean page got truncated */
2620 ret = -EFAULT;
2621 goto out_unlock;
2622 }
2623
2624 end = folio_size(folio);
2625 /* folio is wholly or partially inside EOF */
2626 if (folio_pos(folio) + end > size)
2627 end = size - folio_pos(folio);
2628
2629 ret = __block_write_begin_int(folio, 0, end, get_block, NULL);
2630 if (unlikely(ret))
2631 goto out_unlock;
2632
2633 __block_commit_write(folio, 0, end);
2634
2635 folio_mark_dirty(folio);
2636 folio_wait_stable(folio);
2637 return 0;
2638out_unlock:
2639 folio_unlock(folio);
2640 return ret;
2641}
2642EXPORT_SYMBOL(block_page_mkwrite);
2643
2644int block_truncate_page(struct address_space *mapping,
2645 loff_t from, get_block_t *get_block)
2646{
2647 pgoff_t index = from >> PAGE_SHIFT;
2648 unsigned blocksize;
2649 sector_t iblock;
2650 size_t offset, length, pos;
2651 struct inode *inode = mapping->host;
2652 struct folio *folio;
2653 struct buffer_head *bh;
2654 int err = 0;
2655
2656 blocksize = i_blocksize(inode);
2657 length = from & (blocksize - 1);
2658
2659 /* Block boundary? Nothing to do */
2660 if (!length)
2661 return 0;
2662
2663 length = blocksize - length;
2664 iblock = ((loff_t)index * PAGE_SIZE) >> inode->i_blkbits;
2665
2666 folio = filemap_grab_folio(mapping, index);
2667 if (IS_ERR(folio))
2668 return PTR_ERR(folio);
2669
2670 bh = folio_buffers(folio);
2671 if (!bh)
2672 bh = create_empty_buffers(folio, blocksize, 0);
2673
2674 /* Find the buffer that contains "offset" */
2675 offset = offset_in_folio(folio, from);
2676 pos = blocksize;
2677 while (offset >= pos) {
2678 bh = bh->b_this_page;
2679 iblock++;
2680 pos += blocksize;
2681 }
2682
2683 if (!buffer_mapped(bh)) {
2684 WARN_ON(bh->b_size != blocksize);
2685 err = get_block(inode, iblock, bh, 0);
2686 if (err)
2687 goto unlock;
2688 /* unmapped? It's a hole - nothing to do */
2689 if (!buffer_mapped(bh))
2690 goto unlock;
2691 }
2692
2693 /* Ok, it's mapped. Make sure it's up-to-date */
2694 if (folio_test_uptodate(folio))
2695 set_buffer_uptodate(bh);
2696
2697 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2698 err = bh_read(bh, 0);
2699 /* Uhhuh. Read error. Complain and punt. */
2700 if (err < 0)
2701 goto unlock;
2702 }
2703
2704 folio_zero_range(folio, offset, length);
2705 mark_buffer_dirty(bh);
2706
2707unlock:
2708 folio_unlock(folio);
2709 folio_put(folio);
2710
2711 return err;
2712}
2713EXPORT_SYMBOL(block_truncate_page);
2714
2715/*
2716 * The generic ->writepage function for buffer-backed address_spaces
2717 */
2718int block_write_full_folio(struct folio *folio, struct writeback_control *wbc,
2719 void *get_block)
2720{
2721 struct inode * const inode = folio->mapping->host;
2722 loff_t i_size = i_size_read(inode);
2723
2724 /* Is the folio fully inside i_size? */
2725 if (folio_pos(folio) + folio_size(folio) <= i_size)
2726 return __block_write_full_folio(inode, folio, get_block, wbc);
2727
2728 /* Is the folio fully outside i_size? (truncate in progress) */
2729 if (folio_pos(folio) >= i_size) {
2730 folio_unlock(folio);
2731 return 0; /* don't care */
2732 }
2733
2734 /*
2735 * The folio straddles i_size. It must be zeroed out on each and every
2736 * writepage invocation because it may be mmapped. "A file is mapped
2737 * in multiples of the page size. For a file that is not a multiple of
2738 * the page size, the remaining memory is zeroed when mapped, and
2739 * writes to that region are not written out to the file."
2740 */
2741 folio_zero_segment(folio, offset_in_folio(folio, i_size),
2742 folio_size(folio));
2743 return __block_write_full_folio(inode, folio, get_block, wbc);
2744}
2745
2746sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2747 get_block_t *get_block)
2748{
2749 struct inode *inode = mapping->host;
2750 struct buffer_head tmp = {
2751 .b_size = i_blocksize(inode),
2752 };
2753
2754 get_block(inode, block, &tmp, 0);
2755 return tmp.b_blocknr;
2756}
2757EXPORT_SYMBOL(generic_block_bmap);
2758
2759static void end_bio_bh_io_sync(struct bio *bio)
2760{
2761 struct buffer_head *bh = bio->bi_private;
2762
2763 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2764 set_bit(BH_Quiet, &bh->b_state);
2765
2766 bh->b_end_io(bh, !bio->bi_status);
2767 bio_put(bio);
2768}
2769
2770static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2771 enum rw_hint write_hint,
2772 struct writeback_control *wbc)
2773{
2774 const enum req_op op = opf & REQ_OP_MASK;
2775 struct bio *bio;
2776
2777 BUG_ON(!buffer_locked(bh));
2778 BUG_ON(!buffer_mapped(bh));
2779 BUG_ON(!bh->b_end_io);
2780 BUG_ON(buffer_delay(bh));
2781 BUG_ON(buffer_unwritten(bh));
2782
2783 /*
2784 * Only clear out a write error when rewriting
2785 */
2786 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2787 clear_buffer_write_io_error(bh);
2788
2789 if (buffer_meta(bh))
2790 opf |= REQ_META;
2791 if (buffer_prio(bh))
2792 opf |= REQ_PRIO;
2793
2794 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2795
2796 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2797
2798 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2799 bio->bi_write_hint = write_hint;
2800
2801 bio_add_folio_nofail(bio, bh->b_folio, bh->b_size, bh_offset(bh));
2802
2803 bio->bi_end_io = end_bio_bh_io_sync;
2804 bio->bi_private = bh;
2805
2806 /* Take care of bh's that straddle the end of the device */
2807 guard_bio_eod(bio);
2808
2809 if (wbc) {
2810 wbc_init_bio(wbc, bio);
2811 wbc_account_cgroup_owner(wbc, bh->b_folio, bh->b_size);
2812 }
2813
2814 submit_bio(bio);
2815}
2816
2817void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2818{
2819 submit_bh_wbc(opf, bh, WRITE_LIFE_NOT_SET, NULL);
2820}
2821EXPORT_SYMBOL(submit_bh);
2822
2823void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2824{
2825 lock_buffer(bh);
2826 if (!test_clear_buffer_dirty(bh)) {
2827 unlock_buffer(bh);
2828 return;
2829 }
2830 bh->b_end_io = end_buffer_write_sync;
2831 get_bh(bh);
2832 submit_bh(REQ_OP_WRITE | op_flags, bh);
2833}
2834EXPORT_SYMBOL(write_dirty_buffer);
2835
2836/*
2837 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2838 * and then start new I/O and then wait upon it. The caller must have a ref on
2839 * the buffer_head.
2840 */
2841int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2842{
2843 WARN_ON(atomic_read(&bh->b_count) < 1);
2844 lock_buffer(bh);
2845 if (test_clear_buffer_dirty(bh)) {
2846 /*
2847 * The bh should be mapped, but it might not be if the
2848 * device was hot-removed. Not much we can do but fail the I/O.
2849 */
2850 if (!buffer_mapped(bh)) {
2851 unlock_buffer(bh);
2852 return -EIO;
2853 }
2854
2855 get_bh(bh);
2856 bh->b_end_io = end_buffer_write_sync;
2857 submit_bh(REQ_OP_WRITE | op_flags, bh);
2858 wait_on_buffer(bh);
2859 if (!buffer_uptodate(bh))
2860 return -EIO;
2861 } else {
2862 unlock_buffer(bh);
2863 }
2864 return 0;
2865}
2866EXPORT_SYMBOL(__sync_dirty_buffer);
2867
2868int sync_dirty_buffer(struct buffer_head *bh)
2869{
2870 return __sync_dirty_buffer(bh, REQ_SYNC);
2871}
2872EXPORT_SYMBOL(sync_dirty_buffer);
2873
2874static inline int buffer_busy(struct buffer_head *bh)
2875{
2876 return atomic_read(&bh->b_count) |
2877 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2878}
2879
2880static bool
2881drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2882{
2883 struct buffer_head *head = folio_buffers(folio);
2884 struct buffer_head *bh;
2885
2886 bh = head;
2887 do {
2888 if (buffer_busy(bh))
2889 goto failed;
2890 bh = bh->b_this_page;
2891 } while (bh != head);
2892
2893 do {
2894 struct buffer_head *next = bh->b_this_page;
2895
2896 if (bh->b_assoc_map)
2897 __remove_assoc_queue(bh);
2898 bh = next;
2899 } while (bh != head);
2900 *buffers_to_free = head;
2901 folio_detach_private(folio);
2902 return true;
2903failed:
2904 return false;
2905}
2906
2907/**
2908 * try_to_free_buffers - Release buffers attached to this folio.
2909 * @folio: The folio.
2910 *
2911 * If any buffers are in use (dirty, under writeback, elevated refcount),
2912 * no buffers will be freed.
2913 *
2914 * If the folio is dirty but all the buffers are clean then we need to
2915 * be sure to mark the folio clean as well. This is because the folio
2916 * may be against a block device, and a later reattachment of buffers
2917 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2918 * filesystem data on the same device.
2919 *
2920 * The same applies to regular filesystem folios: if all the buffers are
2921 * clean then we set the folio clean and proceed. To do that, we require
2922 * total exclusion from block_dirty_folio(). That is obtained with
2923 * i_private_lock.
2924 *
2925 * Exclusion against try_to_free_buffers may be obtained by either
2926 * locking the folio or by holding its mapping's i_private_lock.
2927 *
2928 * Context: Process context. @folio must be locked. Will not sleep.
2929 * Return: true if all buffers attached to this folio were freed.
2930 */
2931bool try_to_free_buffers(struct folio *folio)
2932{
2933 struct address_space * const mapping = folio->mapping;
2934 struct buffer_head *buffers_to_free = NULL;
2935 bool ret = 0;
2936
2937 BUG_ON(!folio_test_locked(folio));
2938 if (folio_test_writeback(folio))
2939 return false;
2940
2941 if (mapping == NULL) { /* can this still happen? */
2942 ret = drop_buffers(folio, &buffers_to_free);
2943 goto out;
2944 }
2945
2946 spin_lock(&mapping->i_private_lock);
2947 ret = drop_buffers(folio, &buffers_to_free);
2948
2949 /*
2950 * If the filesystem writes its buffers by hand (eg ext3)
2951 * then we can have clean buffers against a dirty folio. We
2952 * clean the folio here; otherwise the VM will never notice
2953 * that the filesystem did any IO at all.
2954 *
2955 * Also, during truncate, discard_buffer will have marked all
2956 * the folio's buffers clean. We discover that here and clean
2957 * the folio also.
2958 *
2959 * i_private_lock must be held over this entire operation in order
2960 * to synchronise against block_dirty_folio and prevent the
2961 * dirty bit from being lost.
2962 */
2963 if (ret)
2964 folio_cancel_dirty(folio);
2965 spin_unlock(&mapping->i_private_lock);
2966out:
2967 if (buffers_to_free) {
2968 struct buffer_head *bh = buffers_to_free;
2969
2970 do {
2971 struct buffer_head *next = bh->b_this_page;
2972 free_buffer_head(bh);
2973 bh = next;
2974 } while (bh != buffers_to_free);
2975 }
2976 return ret;
2977}
2978EXPORT_SYMBOL(try_to_free_buffers);
2979
2980/*
2981 * Buffer-head allocation
2982 */
2983static struct kmem_cache *bh_cachep __ro_after_init;
2984
2985/*
2986 * Once the number of bh's in the machine exceeds this level, we start
2987 * stripping them in writeback.
2988 */
2989static unsigned long max_buffer_heads __ro_after_init;
2990
2991int buffer_heads_over_limit;
2992
2993struct bh_accounting {
2994 int nr; /* Number of live bh's */
2995 int ratelimit; /* Limit cacheline bouncing */
2996};
2997
2998static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2999
3000static void recalc_bh_state(void)
3001{
3002 int i;
3003 int tot = 0;
3004
3005 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3006 return;
3007 __this_cpu_write(bh_accounting.ratelimit, 0);
3008 for_each_online_cpu(i)
3009 tot += per_cpu(bh_accounting, i).nr;
3010 buffer_heads_over_limit = (tot > max_buffer_heads);
3011}
3012
3013struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3014{
3015 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3016 if (ret) {
3017 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3018 spin_lock_init(&ret->b_uptodate_lock);
3019 preempt_disable();
3020 __this_cpu_inc(bh_accounting.nr);
3021 recalc_bh_state();
3022 preempt_enable();
3023 }
3024 return ret;
3025}
3026EXPORT_SYMBOL(alloc_buffer_head);
3027
3028void free_buffer_head(struct buffer_head *bh)
3029{
3030 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3031 kmem_cache_free(bh_cachep, bh);
3032 preempt_disable();
3033 __this_cpu_dec(bh_accounting.nr);
3034 recalc_bh_state();
3035 preempt_enable();
3036}
3037EXPORT_SYMBOL(free_buffer_head);
3038
3039static int buffer_exit_cpu_dead(unsigned int cpu)
3040{
3041 int i;
3042 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3043
3044 for (i = 0; i < BH_LRU_SIZE; i++) {
3045 brelse(b->bhs[i]);
3046 b->bhs[i] = NULL;
3047 }
3048 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3049 per_cpu(bh_accounting, cpu).nr = 0;
3050 return 0;
3051}
3052
3053/**
3054 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3055 * @bh: struct buffer_head
3056 *
3057 * Return true if the buffer is up-to-date and false,
3058 * with the buffer locked, if not.
3059 */
3060int bh_uptodate_or_lock(struct buffer_head *bh)
3061{
3062 if (!buffer_uptodate(bh)) {
3063 lock_buffer(bh);
3064 if (!buffer_uptodate(bh))
3065 return 0;
3066 unlock_buffer(bh);
3067 }
3068 return 1;
3069}
3070EXPORT_SYMBOL(bh_uptodate_or_lock);
3071
3072/**
3073 * __bh_read - Submit read for a locked buffer
3074 * @bh: struct buffer_head
3075 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3076 * @wait: wait until reading finish
3077 *
3078 * Returns zero on success or don't wait, and -EIO on error.
3079 */
3080int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
3081{
3082 int ret = 0;
3083
3084 BUG_ON(!buffer_locked(bh));
3085
3086 get_bh(bh);
3087 bh->b_end_io = end_buffer_read_sync;
3088 submit_bh(REQ_OP_READ | op_flags, bh);
3089 if (wait) {
3090 wait_on_buffer(bh);
3091 if (!buffer_uptodate(bh))
3092 ret = -EIO;
3093 }
3094 return ret;
3095}
3096EXPORT_SYMBOL(__bh_read);
3097
3098/**
3099 * __bh_read_batch - Submit read for a batch of unlocked buffers
3100 * @nr: entry number of the buffer batch
3101 * @bhs: a batch of struct buffer_head
3102 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3103 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
3104 * buffer that cannot lock.
3105 *
3106 * Returns zero on success or don't wait, and -EIO on error.
3107 */
3108void __bh_read_batch(int nr, struct buffer_head *bhs[],
3109 blk_opf_t op_flags, bool force_lock)
3110{
3111 int i;
3112
3113 for (i = 0; i < nr; i++) {
3114 struct buffer_head *bh = bhs[i];
3115
3116 if (buffer_uptodate(bh))
3117 continue;
3118
3119 if (force_lock)
3120 lock_buffer(bh);
3121 else
3122 if (!trylock_buffer(bh))
3123 continue;
3124
3125 if (buffer_uptodate(bh)) {
3126 unlock_buffer(bh);
3127 continue;
3128 }
3129
3130 bh->b_end_io = end_buffer_read_sync;
3131 get_bh(bh);
3132 submit_bh(REQ_OP_READ | op_flags, bh);
3133 }
3134}
3135EXPORT_SYMBOL(__bh_read_batch);
3136
3137void __init buffer_init(void)
3138{
3139 unsigned long nrpages;
3140 int ret;
3141
3142 bh_cachep = KMEM_CACHE(buffer_head,
3143 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC);
3144 /*
3145 * Limit the bh occupancy to 10% of ZONE_NORMAL
3146 */
3147 nrpages = (nr_free_buffer_pages() * 10) / 100;
3148 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3149 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3150 NULL, buffer_exit_cpu_dead);
3151 WARN_ON(ret < 0);
3152}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/buffer.c
4 *
5 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
6 */
7
8/*
9 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10 *
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
13 *
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
16 *
17 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 *
19 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
20 */
21
22#include <linux/kernel.h>
23#include <linux/sched/signal.h>
24#include <linux/syscalls.h>
25#include <linux/fs.h>
26#include <linux/iomap.h>
27#include <linux/mm.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
52#include "internal.h"
53
54static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
55static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
56 struct writeback_control *wbc);
57
58#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
59
60inline void touch_buffer(struct buffer_head *bh)
61{
62 trace_block_touch_buffer(bh);
63 mark_page_accessed(bh->b_page);
64}
65EXPORT_SYMBOL(touch_buffer);
66
67void __lock_buffer(struct buffer_head *bh)
68{
69 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
70}
71EXPORT_SYMBOL(__lock_buffer);
72
73void unlock_buffer(struct buffer_head *bh)
74{
75 clear_bit_unlock(BH_Lock, &bh->b_state);
76 smp_mb__after_atomic();
77 wake_up_bit(&bh->b_state, BH_Lock);
78}
79EXPORT_SYMBOL(unlock_buffer);
80
81/*
82 * Returns if the folio has dirty or writeback buffers. If all the buffers
83 * are unlocked and clean then the folio_test_dirty information is stale. If
84 * any of the buffers are locked, it is assumed they are locked for IO.
85 */
86void buffer_check_dirty_writeback(struct folio *folio,
87 bool *dirty, bool *writeback)
88{
89 struct buffer_head *head, *bh;
90 *dirty = false;
91 *writeback = false;
92
93 BUG_ON(!folio_test_locked(folio));
94
95 head = folio_buffers(folio);
96 if (!head)
97 return;
98
99 if (folio_test_writeback(folio))
100 *writeback = true;
101
102 bh = head;
103 do {
104 if (buffer_locked(bh))
105 *writeback = true;
106
107 if (buffer_dirty(bh))
108 *dirty = true;
109
110 bh = bh->b_this_page;
111 } while (bh != head);
112}
113EXPORT_SYMBOL(buffer_check_dirty_writeback);
114
115/*
116 * Block until a buffer comes unlocked. This doesn't stop it
117 * from becoming locked again - you have to lock it yourself
118 * if you want to preserve its state.
119 */
120void __wait_on_buffer(struct buffer_head * bh)
121{
122 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
123}
124EXPORT_SYMBOL(__wait_on_buffer);
125
126static void buffer_io_error(struct buffer_head *bh, char *msg)
127{
128 if (!test_bit(BH_Quiet, &bh->b_state))
129 printk_ratelimited(KERN_ERR
130 "Buffer I/O error on dev %pg, logical block %llu%s\n",
131 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
132}
133
134/*
135 * End-of-IO handler helper function which does not touch the bh after
136 * unlocking it.
137 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
138 * a race there is benign: unlock_buffer() only use the bh's address for
139 * hashing after unlocking the buffer, so it doesn't actually touch the bh
140 * itself.
141 */
142static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
143{
144 if (uptodate) {
145 set_buffer_uptodate(bh);
146 } else {
147 /* This happens, due to failed read-ahead attempts. */
148 clear_buffer_uptodate(bh);
149 }
150 unlock_buffer(bh);
151}
152
153/*
154 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
155 * unlock the buffer.
156 */
157void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
158{
159 __end_buffer_read_notouch(bh, uptodate);
160 put_bh(bh);
161}
162EXPORT_SYMBOL(end_buffer_read_sync);
163
164void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
165{
166 if (uptodate) {
167 set_buffer_uptodate(bh);
168 } else {
169 buffer_io_error(bh, ", lost sync page write");
170 mark_buffer_write_io_error(bh);
171 clear_buffer_uptodate(bh);
172 }
173 unlock_buffer(bh);
174 put_bh(bh);
175}
176EXPORT_SYMBOL(end_buffer_write_sync);
177
178/*
179 * Various filesystems appear to want __find_get_block to be non-blocking.
180 * But it's the page lock which protects the buffers. To get around this,
181 * we get exclusion from try_to_free_buffers with the blockdev mapping's
182 * private_lock.
183 *
184 * Hack idea: for the blockdev mapping, private_lock contention
185 * may be quite high. This code could TryLock the page, and if that
186 * succeeds, there is no need to take private_lock.
187 */
188static struct buffer_head *
189__find_get_block_slow(struct block_device *bdev, sector_t block)
190{
191 struct inode *bd_inode = bdev->bd_inode;
192 struct address_space *bd_mapping = bd_inode->i_mapping;
193 struct buffer_head *ret = NULL;
194 pgoff_t index;
195 struct buffer_head *bh;
196 struct buffer_head *head;
197 struct page *page;
198 int all_mapped = 1;
199 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
200
201 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
202 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
203 if (!page)
204 goto out;
205
206 spin_lock(&bd_mapping->private_lock);
207 if (!page_has_buffers(page))
208 goto out_unlock;
209 head = page_buffers(page);
210 bh = head;
211 do {
212 if (!buffer_mapped(bh))
213 all_mapped = 0;
214 else if (bh->b_blocknr == block) {
215 ret = bh;
216 get_bh(bh);
217 goto out_unlock;
218 }
219 bh = bh->b_this_page;
220 } while (bh != head);
221
222 /* we might be here because some of the buffers on this page are
223 * not mapped. This is due to various races between
224 * file io on the block device and getblk. It gets dealt with
225 * elsewhere, don't buffer_error if we had some unmapped buffers
226 */
227 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
228 if (all_mapped && __ratelimit(&last_warned)) {
229 printk("__find_get_block_slow() failed. block=%llu, "
230 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
231 "device %pg blocksize: %d\n",
232 (unsigned long long)block,
233 (unsigned long long)bh->b_blocknr,
234 bh->b_state, bh->b_size, bdev,
235 1 << bd_inode->i_blkbits);
236 }
237out_unlock:
238 spin_unlock(&bd_mapping->private_lock);
239 put_page(page);
240out:
241 return ret;
242}
243
244static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
245{
246 unsigned long flags;
247 struct buffer_head *first;
248 struct buffer_head *tmp;
249 struct page *page;
250 int page_uptodate = 1;
251
252 BUG_ON(!buffer_async_read(bh));
253
254 page = bh->b_page;
255 if (uptodate) {
256 set_buffer_uptodate(bh);
257 } else {
258 clear_buffer_uptodate(bh);
259 buffer_io_error(bh, ", async page read");
260 SetPageError(page);
261 }
262
263 /*
264 * Be _very_ careful from here on. Bad things can happen if
265 * two buffer heads end IO at almost the same time and both
266 * decide that the page is now completely done.
267 */
268 first = page_buffers(page);
269 spin_lock_irqsave(&first->b_uptodate_lock, flags);
270 clear_buffer_async_read(bh);
271 unlock_buffer(bh);
272 tmp = bh;
273 do {
274 if (!buffer_uptodate(tmp))
275 page_uptodate = 0;
276 if (buffer_async_read(tmp)) {
277 BUG_ON(!buffer_locked(tmp));
278 goto still_busy;
279 }
280 tmp = tmp->b_this_page;
281 } while (tmp != bh);
282 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
283
284 /*
285 * If all of the buffers are uptodate then we can set the page
286 * uptodate.
287 */
288 if (page_uptodate)
289 SetPageUptodate(page);
290 unlock_page(page);
291 return;
292
293still_busy:
294 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
295 return;
296}
297
298struct decrypt_bh_ctx {
299 struct work_struct work;
300 struct buffer_head *bh;
301};
302
303static void decrypt_bh(struct work_struct *work)
304{
305 struct decrypt_bh_ctx *ctx =
306 container_of(work, struct decrypt_bh_ctx, work);
307 struct buffer_head *bh = ctx->bh;
308 int err;
309
310 err = fscrypt_decrypt_pagecache_blocks(bh->b_page, bh->b_size,
311 bh_offset(bh));
312 end_buffer_async_read(bh, err == 0);
313 kfree(ctx);
314}
315
316/*
317 * I/O completion handler for block_read_full_folio() - pages
318 * which come unlocked at the end of I/O.
319 */
320static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
321{
322 /* Decrypt if needed */
323 if (uptodate &&
324 fscrypt_inode_uses_fs_layer_crypto(bh->b_page->mapping->host)) {
325 struct decrypt_bh_ctx *ctx = kmalloc(sizeof(*ctx), GFP_ATOMIC);
326
327 if (ctx) {
328 INIT_WORK(&ctx->work, decrypt_bh);
329 ctx->bh = bh;
330 fscrypt_enqueue_decrypt_work(&ctx->work);
331 return;
332 }
333 uptodate = 0;
334 }
335 end_buffer_async_read(bh, uptodate);
336}
337
338/*
339 * Completion handler for block_write_full_page() - pages which are unlocked
340 * during I/O, and which have PageWriteback cleared upon I/O completion.
341 */
342void end_buffer_async_write(struct buffer_head *bh, int uptodate)
343{
344 unsigned long flags;
345 struct buffer_head *first;
346 struct buffer_head *tmp;
347 struct page *page;
348
349 BUG_ON(!buffer_async_write(bh));
350
351 page = bh->b_page;
352 if (uptodate) {
353 set_buffer_uptodate(bh);
354 } else {
355 buffer_io_error(bh, ", lost async page write");
356 mark_buffer_write_io_error(bh);
357 clear_buffer_uptodate(bh);
358 SetPageError(page);
359 }
360
361 first = page_buffers(page);
362 spin_lock_irqsave(&first->b_uptodate_lock, flags);
363
364 clear_buffer_async_write(bh);
365 unlock_buffer(bh);
366 tmp = bh->b_this_page;
367 while (tmp != bh) {
368 if (buffer_async_write(tmp)) {
369 BUG_ON(!buffer_locked(tmp));
370 goto still_busy;
371 }
372 tmp = tmp->b_this_page;
373 }
374 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
375 end_page_writeback(page);
376 return;
377
378still_busy:
379 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
380 return;
381}
382EXPORT_SYMBOL(end_buffer_async_write);
383
384/*
385 * If a page's buffers are under async readin (end_buffer_async_read
386 * completion) then there is a possibility that another thread of
387 * control could lock one of the buffers after it has completed
388 * but while some of the other buffers have not completed. This
389 * locked buffer would confuse end_buffer_async_read() into not unlocking
390 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
391 * that this buffer is not under async I/O.
392 *
393 * The page comes unlocked when it has no locked buffer_async buffers
394 * left.
395 *
396 * PageLocked prevents anyone starting new async I/O reads any of
397 * the buffers.
398 *
399 * PageWriteback is used to prevent simultaneous writeout of the same
400 * page.
401 *
402 * PageLocked prevents anyone from starting writeback of a page which is
403 * under read I/O (PageWriteback is only ever set against a locked page).
404 */
405static void mark_buffer_async_read(struct buffer_head *bh)
406{
407 bh->b_end_io = end_buffer_async_read_io;
408 set_buffer_async_read(bh);
409}
410
411static void mark_buffer_async_write_endio(struct buffer_head *bh,
412 bh_end_io_t *handler)
413{
414 bh->b_end_io = handler;
415 set_buffer_async_write(bh);
416}
417
418void mark_buffer_async_write(struct buffer_head *bh)
419{
420 mark_buffer_async_write_endio(bh, end_buffer_async_write);
421}
422EXPORT_SYMBOL(mark_buffer_async_write);
423
424
425/*
426 * fs/buffer.c contains helper functions for buffer-backed address space's
427 * fsync functions. A common requirement for buffer-based filesystems is
428 * that certain data from the backing blockdev needs to be written out for
429 * a successful fsync(). For example, ext2 indirect blocks need to be
430 * written back and waited upon before fsync() returns.
431 *
432 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
433 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
434 * management of a list of dependent buffers at ->i_mapping->private_list.
435 *
436 * Locking is a little subtle: try_to_free_buffers() will remove buffers
437 * from their controlling inode's queue when they are being freed. But
438 * try_to_free_buffers() will be operating against the *blockdev* mapping
439 * at the time, not against the S_ISREG file which depends on those buffers.
440 * So the locking for private_list is via the private_lock in the address_space
441 * which backs the buffers. Which is different from the address_space
442 * against which the buffers are listed. So for a particular address_space,
443 * mapping->private_lock does *not* protect mapping->private_list! In fact,
444 * mapping->private_list will always be protected by the backing blockdev's
445 * ->private_lock.
446 *
447 * Which introduces a requirement: all buffers on an address_space's
448 * ->private_list must be from the same address_space: the blockdev's.
449 *
450 * address_spaces which do not place buffers at ->private_list via these
451 * utility functions are free to use private_lock and private_list for
452 * whatever they want. The only requirement is that list_empty(private_list)
453 * be true at clear_inode() time.
454 *
455 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
456 * filesystems should do that. invalidate_inode_buffers() should just go
457 * BUG_ON(!list_empty).
458 *
459 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
460 * take an address_space, not an inode. And it should be called
461 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
462 * queued up.
463 *
464 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
465 * list if it is already on a list. Because if the buffer is on a list,
466 * it *must* already be on the right one. If not, the filesystem is being
467 * silly. This will save a ton of locking. But first we have to ensure
468 * that buffers are taken *off* the old inode's list when they are freed
469 * (presumably in truncate). That requires careful auditing of all
470 * filesystems (do it inside bforget()). It could also be done by bringing
471 * b_inode back.
472 */
473
474/*
475 * The buffer's backing address_space's private_lock must be held
476 */
477static void __remove_assoc_queue(struct buffer_head *bh)
478{
479 list_del_init(&bh->b_assoc_buffers);
480 WARN_ON(!bh->b_assoc_map);
481 bh->b_assoc_map = NULL;
482}
483
484int inode_has_buffers(struct inode *inode)
485{
486 return !list_empty(&inode->i_data.private_list);
487}
488
489/*
490 * osync is designed to support O_SYNC io. It waits synchronously for
491 * all already-submitted IO to complete, but does not queue any new
492 * writes to the disk.
493 *
494 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
495 * as you dirty the buffers, and then use osync_inode_buffers to wait for
496 * completion. Any other dirty buffers which are not yet queued for
497 * write will not be flushed to disk by the osync.
498 */
499static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
500{
501 struct buffer_head *bh;
502 struct list_head *p;
503 int err = 0;
504
505 spin_lock(lock);
506repeat:
507 list_for_each_prev(p, list) {
508 bh = BH_ENTRY(p);
509 if (buffer_locked(bh)) {
510 get_bh(bh);
511 spin_unlock(lock);
512 wait_on_buffer(bh);
513 if (!buffer_uptodate(bh))
514 err = -EIO;
515 brelse(bh);
516 spin_lock(lock);
517 goto repeat;
518 }
519 }
520 spin_unlock(lock);
521 return err;
522}
523
524void emergency_thaw_bdev(struct super_block *sb)
525{
526 while (sb->s_bdev && !thaw_bdev(sb->s_bdev))
527 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
528}
529
530/**
531 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
532 * @mapping: the mapping which wants those buffers written
533 *
534 * Starts I/O against the buffers at mapping->private_list, and waits upon
535 * that I/O.
536 *
537 * Basically, this is a convenience function for fsync().
538 * @mapping is a file or directory which needs those buffers to be written for
539 * a successful fsync().
540 */
541int sync_mapping_buffers(struct address_space *mapping)
542{
543 struct address_space *buffer_mapping = mapping->private_data;
544
545 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
546 return 0;
547
548 return fsync_buffers_list(&buffer_mapping->private_lock,
549 &mapping->private_list);
550}
551EXPORT_SYMBOL(sync_mapping_buffers);
552
553/*
554 * Called when we've recently written block `bblock', and it is known that
555 * `bblock' was for a buffer_boundary() buffer. This means that the block at
556 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
557 * dirty, schedule it for IO. So that indirects merge nicely with their data.
558 */
559void write_boundary_block(struct block_device *bdev,
560 sector_t bblock, unsigned blocksize)
561{
562 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
563 if (bh) {
564 if (buffer_dirty(bh))
565 write_dirty_buffer(bh, 0);
566 put_bh(bh);
567 }
568}
569
570void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
571{
572 struct address_space *mapping = inode->i_mapping;
573 struct address_space *buffer_mapping = bh->b_page->mapping;
574
575 mark_buffer_dirty(bh);
576 if (!mapping->private_data) {
577 mapping->private_data = buffer_mapping;
578 } else {
579 BUG_ON(mapping->private_data != buffer_mapping);
580 }
581 if (!bh->b_assoc_map) {
582 spin_lock(&buffer_mapping->private_lock);
583 list_move_tail(&bh->b_assoc_buffers,
584 &mapping->private_list);
585 bh->b_assoc_map = mapping;
586 spin_unlock(&buffer_mapping->private_lock);
587 }
588}
589EXPORT_SYMBOL(mark_buffer_dirty_inode);
590
591/*
592 * Add a page to the dirty page list.
593 *
594 * It is a sad fact of life that this function is called from several places
595 * deeply under spinlocking. It may not sleep.
596 *
597 * If the page has buffers, the uptodate buffers are set dirty, to preserve
598 * dirty-state coherency between the page and the buffers. It the page does
599 * not have buffers then when they are later attached they will all be set
600 * dirty.
601 *
602 * The buffers are dirtied before the page is dirtied. There's a small race
603 * window in which a writepage caller may see the page cleanness but not the
604 * buffer dirtiness. That's fine. If this code were to set the page dirty
605 * before the buffers, a concurrent writepage caller could clear the page dirty
606 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
607 * page on the dirty page list.
608 *
609 * We use private_lock to lock against try_to_free_buffers while using the
610 * page's buffer list. Also use this to protect against clean buffers being
611 * added to the page after it was set dirty.
612 *
613 * FIXME: may need to call ->reservepage here as well. That's rather up to the
614 * address_space though.
615 */
616bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
617{
618 struct buffer_head *head;
619 bool newly_dirty;
620
621 spin_lock(&mapping->private_lock);
622 head = folio_buffers(folio);
623 if (head) {
624 struct buffer_head *bh = head;
625
626 do {
627 set_buffer_dirty(bh);
628 bh = bh->b_this_page;
629 } while (bh != head);
630 }
631 /*
632 * Lock out page's memcg migration to keep PageDirty
633 * synchronized with per-memcg dirty page counters.
634 */
635 folio_memcg_lock(folio);
636 newly_dirty = !folio_test_set_dirty(folio);
637 spin_unlock(&mapping->private_lock);
638
639 if (newly_dirty)
640 __folio_mark_dirty(folio, mapping, 1);
641
642 folio_memcg_unlock(folio);
643
644 if (newly_dirty)
645 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
646
647 return newly_dirty;
648}
649EXPORT_SYMBOL(block_dirty_folio);
650
651/*
652 * Write out and wait upon a list of buffers.
653 *
654 * We have conflicting pressures: we want to make sure that all
655 * initially dirty buffers get waited on, but that any subsequently
656 * dirtied buffers don't. After all, we don't want fsync to last
657 * forever if somebody is actively writing to the file.
658 *
659 * Do this in two main stages: first we copy dirty buffers to a
660 * temporary inode list, queueing the writes as we go. Then we clean
661 * up, waiting for those writes to complete.
662 *
663 * During this second stage, any subsequent updates to the file may end
664 * up refiling the buffer on the original inode's dirty list again, so
665 * there is a chance we will end up with a buffer queued for write but
666 * not yet completed on that list. So, as a final cleanup we go through
667 * the osync code to catch these locked, dirty buffers without requeuing
668 * any newly dirty buffers for write.
669 */
670static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
671{
672 struct buffer_head *bh;
673 struct list_head tmp;
674 struct address_space *mapping;
675 int err = 0, err2;
676 struct blk_plug plug;
677
678 INIT_LIST_HEAD(&tmp);
679 blk_start_plug(&plug);
680
681 spin_lock(lock);
682 while (!list_empty(list)) {
683 bh = BH_ENTRY(list->next);
684 mapping = bh->b_assoc_map;
685 __remove_assoc_queue(bh);
686 /* Avoid race with mark_buffer_dirty_inode() which does
687 * a lockless check and we rely on seeing the dirty bit */
688 smp_mb();
689 if (buffer_dirty(bh) || buffer_locked(bh)) {
690 list_add(&bh->b_assoc_buffers, &tmp);
691 bh->b_assoc_map = mapping;
692 if (buffer_dirty(bh)) {
693 get_bh(bh);
694 spin_unlock(lock);
695 /*
696 * Ensure any pending I/O completes so that
697 * write_dirty_buffer() actually writes the
698 * current contents - it is a noop if I/O is
699 * still in flight on potentially older
700 * contents.
701 */
702 write_dirty_buffer(bh, REQ_SYNC);
703
704 /*
705 * Kick off IO for the previous mapping. Note
706 * that we will not run the very last mapping,
707 * wait_on_buffer() will do that for us
708 * through sync_buffer().
709 */
710 brelse(bh);
711 spin_lock(lock);
712 }
713 }
714 }
715
716 spin_unlock(lock);
717 blk_finish_plug(&plug);
718 spin_lock(lock);
719
720 while (!list_empty(&tmp)) {
721 bh = BH_ENTRY(tmp.prev);
722 get_bh(bh);
723 mapping = bh->b_assoc_map;
724 __remove_assoc_queue(bh);
725 /* Avoid race with mark_buffer_dirty_inode() which does
726 * a lockless check and we rely on seeing the dirty bit */
727 smp_mb();
728 if (buffer_dirty(bh)) {
729 list_add(&bh->b_assoc_buffers,
730 &mapping->private_list);
731 bh->b_assoc_map = mapping;
732 }
733 spin_unlock(lock);
734 wait_on_buffer(bh);
735 if (!buffer_uptodate(bh))
736 err = -EIO;
737 brelse(bh);
738 spin_lock(lock);
739 }
740
741 spin_unlock(lock);
742 err2 = osync_buffers_list(lock, list);
743 if (err)
744 return err;
745 else
746 return err2;
747}
748
749/*
750 * Invalidate any and all dirty buffers on a given inode. We are
751 * probably unmounting the fs, but that doesn't mean we have already
752 * done a sync(). Just drop the buffers from the inode list.
753 *
754 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
755 * assumes that all the buffers are against the blockdev. Not true
756 * for reiserfs.
757 */
758void invalidate_inode_buffers(struct inode *inode)
759{
760 if (inode_has_buffers(inode)) {
761 struct address_space *mapping = &inode->i_data;
762 struct list_head *list = &mapping->private_list;
763 struct address_space *buffer_mapping = mapping->private_data;
764
765 spin_lock(&buffer_mapping->private_lock);
766 while (!list_empty(list))
767 __remove_assoc_queue(BH_ENTRY(list->next));
768 spin_unlock(&buffer_mapping->private_lock);
769 }
770}
771EXPORT_SYMBOL(invalidate_inode_buffers);
772
773/*
774 * Remove any clean buffers from the inode's buffer list. This is called
775 * when we're trying to free the inode itself. Those buffers can pin it.
776 *
777 * Returns true if all buffers were removed.
778 */
779int remove_inode_buffers(struct inode *inode)
780{
781 int ret = 1;
782
783 if (inode_has_buffers(inode)) {
784 struct address_space *mapping = &inode->i_data;
785 struct list_head *list = &mapping->private_list;
786 struct address_space *buffer_mapping = mapping->private_data;
787
788 spin_lock(&buffer_mapping->private_lock);
789 while (!list_empty(list)) {
790 struct buffer_head *bh = BH_ENTRY(list->next);
791 if (buffer_dirty(bh)) {
792 ret = 0;
793 break;
794 }
795 __remove_assoc_queue(bh);
796 }
797 spin_unlock(&buffer_mapping->private_lock);
798 }
799 return ret;
800}
801
802/*
803 * Create the appropriate buffers when given a page for data area and
804 * the size of each buffer.. Use the bh->b_this_page linked list to
805 * follow the buffers created. Return NULL if unable to create more
806 * buffers.
807 *
808 * The retry flag is used to differentiate async IO (paging, swapping)
809 * which may not fail from ordinary buffer allocations.
810 */
811struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
812 bool retry)
813{
814 struct buffer_head *bh, *head;
815 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
816 long offset;
817 struct mem_cgroup *memcg, *old_memcg;
818
819 if (retry)
820 gfp |= __GFP_NOFAIL;
821
822 /* The page lock pins the memcg */
823 memcg = page_memcg(page);
824 old_memcg = set_active_memcg(memcg);
825
826 head = NULL;
827 offset = PAGE_SIZE;
828 while ((offset -= size) >= 0) {
829 bh = alloc_buffer_head(gfp);
830 if (!bh)
831 goto no_grow;
832
833 bh->b_this_page = head;
834 bh->b_blocknr = -1;
835 head = bh;
836
837 bh->b_size = size;
838
839 /* Link the buffer to its page */
840 set_bh_page(bh, page, offset);
841 }
842out:
843 set_active_memcg(old_memcg);
844 return head;
845/*
846 * In case anything failed, we just free everything we got.
847 */
848no_grow:
849 if (head) {
850 do {
851 bh = head;
852 head = head->b_this_page;
853 free_buffer_head(bh);
854 } while (head);
855 }
856
857 goto out;
858}
859EXPORT_SYMBOL_GPL(alloc_page_buffers);
860
861static inline void
862link_dev_buffers(struct page *page, struct buffer_head *head)
863{
864 struct buffer_head *bh, *tail;
865
866 bh = head;
867 do {
868 tail = bh;
869 bh = bh->b_this_page;
870 } while (bh);
871 tail->b_this_page = head;
872 attach_page_private(page, head);
873}
874
875static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
876{
877 sector_t retval = ~((sector_t)0);
878 loff_t sz = bdev_nr_bytes(bdev);
879
880 if (sz) {
881 unsigned int sizebits = blksize_bits(size);
882 retval = (sz >> sizebits);
883 }
884 return retval;
885}
886
887/*
888 * Initialise the state of a blockdev page's buffers.
889 */
890static sector_t
891init_page_buffers(struct page *page, struct block_device *bdev,
892 sector_t block, int size)
893{
894 struct buffer_head *head = page_buffers(page);
895 struct buffer_head *bh = head;
896 int uptodate = PageUptodate(page);
897 sector_t end_block = blkdev_max_block(bdev, size);
898
899 do {
900 if (!buffer_mapped(bh)) {
901 bh->b_end_io = NULL;
902 bh->b_private = NULL;
903 bh->b_bdev = bdev;
904 bh->b_blocknr = block;
905 if (uptodate)
906 set_buffer_uptodate(bh);
907 if (block < end_block)
908 set_buffer_mapped(bh);
909 }
910 block++;
911 bh = bh->b_this_page;
912 } while (bh != head);
913
914 /*
915 * Caller needs to validate requested block against end of device.
916 */
917 return end_block;
918}
919
920/*
921 * Create the page-cache page that contains the requested block.
922 *
923 * This is used purely for blockdev mappings.
924 */
925static int
926grow_dev_page(struct block_device *bdev, sector_t block,
927 pgoff_t index, int size, int sizebits, gfp_t gfp)
928{
929 struct inode *inode = bdev->bd_inode;
930 struct page *page;
931 struct buffer_head *bh;
932 sector_t end_block;
933 int ret = 0;
934 gfp_t gfp_mask;
935
936 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
937
938 /*
939 * XXX: __getblk_slow() can not really deal with failure and
940 * will endlessly loop on improvised global reclaim. Prefer
941 * looping in the allocator rather than here, at least that
942 * code knows what it's doing.
943 */
944 gfp_mask |= __GFP_NOFAIL;
945
946 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
947
948 BUG_ON(!PageLocked(page));
949
950 if (page_has_buffers(page)) {
951 bh = page_buffers(page);
952 if (bh->b_size == size) {
953 end_block = init_page_buffers(page, bdev,
954 (sector_t)index << sizebits,
955 size);
956 goto done;
957 }
958 if (!try_to_free_buffers(page_folio(page)))
959 goto failed;
960 }
961
962 /*
963 * Allocate some buffers for this page
964 */
965 bh = alloc_page_buffers(page, size, true);
966
967 /*
968 * Link the page to the buffers and initialise them. Take the
969 * lock to be atomic wrt __find_get_block(), which does not
970 * run under the page lock.
971 */
972 spin_lock(&inode->i_mapping->private_lock);
973 link_dev_buffers(page, bh);
974 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
975 size);
976 spin_unlock(&inode->i_mapping->private_lock);
977done:
978 ret = (block < end_block) ? 1 : -ENXIO;
979failed:
980 unlock_page(page);
981 put_page(page);
982 return ret;
983}
984
985/*
986 * Create buffers for the specified block device block's page. If
987 * that page was dirty, the buffers are set dirty also.
988 */
989static int
990grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
991{
992 pgoff_t index;
993 int sizebits;
994
995 sizebits = PAGE_SHIFT - __ffs(size);
996 index = block >> sizebits;
997
998 /*
999 * Check for a block which wants to lie outside our maximum possible
1000 * pagecache index. (this comparison is done using sector_t types).
1001 */
1002 if (unlikely(index != block >> sizebits)) {
1003 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1004 "device %pg\n",
1005 __func__, (unsigned long long)block,
1006 bdev);
1007 return -EIO;
1008 }
1009
1010 /* Create a page with the proper size buffers.. */
1011 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1012}
1013
1014static struct buffer_head *
1015__getblk_slow(struct block_device *bdev, sector_t block,
1016 unsigned size, gfp_t gfp)
1017{
1018 /* Size must be multiple of hard sectorsize */
1019 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1020 (size < 512 || size > PAGE_SIZE))) {
1021 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1022 size);
1023 printk(KERN_ERR "logical block size: %d\n",
1024 bdev_logical_block_size(bdev));
1025
1026 dump_stack();
1027 return NULL;
1028 }
1029
1030 for (;;) {
1031 struct buffer_head *bh;
1032 int ret;
1033
1034 bh = __find_get_block(bdev, block, size);
1035 if (bh)
1036 return bh;
1037
1038 ret = grow_buffers(bdev, block, size, gfp);
1039 if (ret < 0)
1040 return NULL;
1041 }
1042}
1043
1044/*
1045 * The relationship between dirty buffers and dirty pages:
1046 *
1047 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1048 * the page is tagged dirty in the page cache.
1049 *
1050 * At all times, the dirtiness of the buffers represents the dirtiness of
1051 * subsections of the page. If the page has buffers, the page dirty bit is
1052 * merely a hint about the true dirty state.
1053 *
1054 * When a page is set dirty in its entirety, all its buffers are marked dirty
1055 * (if the page has buffers).
1056 *
1057 * When a buffer is marked dirty, its page is dirtied, but the page's other
1058 * buffers are not.
1059 *
1060 * Also. When blockdev buffers are explicitly read with bread(), they
1061 * individually become uptodate. But their backing page remains not
1062 * uptodate - even if all of its buffers are uptodate. A subsequent
1063 * block_read_full_folio() against that folio will discover all the uptodate
1064 * buffers, will set the folio uptodate and will perform no I/O.
1065 */
1066
1067/**
1068 * mark_buffer_dirty - mark a buffer_head as needing writeout
1069 * @bh: the buffer_head to mark dirty
1070 *
1071 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1072 * its backing page dirty, then tag the page as dirty in the page cache
1073 * and then attach the address_space's inode to its superblock's dirty
1074 * inode list.
1075 *
1076 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1077 * i_pages lock and mapping->host->i_lock.
1078 */
1079void mark_buffer_dirty(struct buffer_head *bh)
1080{
1081 WARN_ON_ONCE(!buffer_uptodate(bh));
1082
1083 trace_block_dirty_buffer(bh);
1084
1085 /*
1086 * Very *carefully* optimize the it-is-already-dirty case.
1087 *
1088 * Don't let the final "is it dirty" escape to before we
1089 * perhaps modified the buffer.
1090 */
1091 if (buffer_dirty(bh)) {
1092 smp_mb();
1093 if (buffer_dirty(bh))
1094 return;
1095 }
1096
1097 if (!test_set_buffer_dirty(bh)) {
1098 struct page *page = bh->b_page;
1099 struct address_space *mapping = NULL;
1100
1101 lock_page_memcg(page);
1102 if (!TestSetPageDirty(page)) {
1103 mapping = page_mapping(page);
1104 if (mapping)
1105 __set_page_dirty(page, mapping, 0);
1106 }
1107 unlock_page_memcg(page);
1108 if (mapping)
1109 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1110 }
1111}
1112EXPORT_SYMBOL(mark_buffer_dirty);
1113
1114void mark_buffer_write_io_error(struct buffer_head *bh)
1115{
1116 struct super_block *sb;
1117
1118 set_buffer_write_io_error(bh);
1119 /* FIXME: do we need to set this in both places? */
1120 if (bh->b_page && bh->b_page->mapping)
1121 mapping_set_error(bh->b_page->mapping, -EIO);
1122 if (bh->b_assoc_map)
1123 mapping_set_error(bh->b_assoc_map, -EIO);
1124 rcu_read_lock();
1125 sb = READ_ONCE(bh->b_bdev->bd_super);
1126 if (sb)
1127 errseq_set(&sb->s_wb_err, -EIO);
1128 rcu_read_unlock();
1129}
1130EXPORT_SYMBOL(mark_buffer_write_io_error);
1131
1132/*
1133 * Decrement a buffer_head's reference count. If all buffers against a page
1134 * have zero reference count, are clean and unlocked, and if the page is clean
1135 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1136 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1137 * a page but it ends up not being freed, and buffers may later be reattached).
1138 */
1139void __brelse(struct buffer_head * buf)
1140{
1141 if (atomic_read(&buf->b_count)) {
1142 put_bh(buf);
1143 return;
1144 }
1145 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1146}
1147EXPORT_SYMBOL(__brelse);
1148
1149/*
1150 * bforget() is like brelse(), except it discards any
1151 * potentially dirty data.
1152 */
1153void __bforget(struct buffer_head *bh)
1154{
1155 clear_buffer_dirty(bh);
1156 if (bh->b_assoc_map) {
1157 struct address_space *buffer_mapping = bh->b_page->mapping;
1158
1159 spin_lock(&buffer_mapping->private_lock);
1160 list_del_init(&bh->b_assoc_buffers);
1161 bh->b_assoc_map = NULL;
1162 spin_unlock(&buffer_mapping->private_lock);
1163 }
1164 __brelse(bh);
1165}
1166EXPORT_SYMBOL(__bforget);
1167
1168static struct buffer_head *__bread_slow(struct buffer_head *bh)
1169{
1170 lock_buffer(bh);
1171 if (buffer_uptodate(bh)) {
1172 unlock_buffer(bh);
1173 return bh;
1174 } else {
1175 get_bh(bh);
1176 bh->b_end_io = end_buffer_read_sync;
1177 submit_bh(REQ_OP_READ, bh);
1178 wait_on_buffer(bh);
1179 if (buffer_uptodate(bh))
1180 return bh;
1181 }
1182 brelse(bh);
1183 return NULL;
1184}
1185
1186/*
1187 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1188 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1189 * refcount elevated by one when they're in an LRU. A buffer can only appear
1190 * once in a particular CPU's LRU. A single buffer can be present in multiple
1191 * CPU's LRUs at the same time.
1192 *
1193 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1194 * sb_find_get_block().
1195 *
1196 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1197 * a local interrupt disable for that.
1198 */
1199
1200#define BH_LRU_SIZE 16
1201
1202struct bh_lru {
1203 struct buffer_head *bhs[BH_LRU_SIZE];
1204};
1205
1206static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1207
1208#ifdef CONFIG_SMP
1209#define bh_lru_lock() local_irq_disable()
1210#define bh_lru_unlock() local_irq_enable()
1211#else
1212#define bh_lru_lock() preempt_disable()
1213#define bh_lru_unlock() preempt_enable()
1214#endif
1215
1216static inline void check_irqs_on(void)
1217{
1218#ifdef irqs_disabled
1219 BUG_ON(irqs_disabled());
1220#endif
1221}
1222
1223/*
1224 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1225 * inserted at the front, and the buffer_head at the back if any is evicted.
1226 * Or, if already in the LRU it is moved to the front.
1227 */
1228static void bh_lru_install(struct buffer_head *bh)
1229{
1230 struct buffer_head *evictee = bh;
1231 struct bh_lru *b;
1232 int i;
1233
1234 check_irqs_on();
1235 bh_lru_lock();
1236
1237 /*
1238 * the refcount of buffer_head in bh_lru prevents dropping the
1239 * attached page(i.e., try_to_free_buffers) so it could cause
1240 * failing page migration.
1241 * Skip putting upcoming bh into bh_lru until migration is done.
1242 */
1243 if (lru_cache_disabled()) {
1244 bh_lru_unlock();
1245 return;
1246 }
1247
1248 b = this_cpu_ptr(&bh_lrus);
1249 for (i = 0; i < BH_LRU_SIZE; i++) {
1250 swap(evictee, b->bhs[i]);
1251 if (evictee == bh) {
1252 bh_lru_unlock();
1253 return;
1254 }
1255 }
1256
1257 get_bh(bh);
1258 bh_lru_unlock();
1259 brelse(evictee);
1260}
1261
1262/*
1263 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1264 */
1265static struct buffer_head *
1266lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1267{
1268 struct buffer_head *ret = NULL;
1269 unsigned int i;
1270
1271 check_irqs_on();
1272 bh_lru_lock();
1273 for (i = 0; i < BH_LRU_SIZE; i++) {
1274 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1275
1276 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1277 bh->b_size == size) {
1278 if (i) {
1279 while (i) {
1280 __this_cpu_write(bh_lrus.bhs[i],
1281 __this_cpu_read(bh_lrus.bhs[i - 1]));
1282 i--;
1283 }
1284 __this_cpu_write(bh_lrus.bhs[0], bh);
1285 }
1286 get_bh(bh);
1287 ret = bh;
1288 break;
1289 }
1290 }
1291 bh_lru_unlock();
1292 return ret;
1293}
1294
1295/*
1296 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1297 * it in the LRU and mark it as accessed. If it is not present then return
1298 * NULL
1299 */
1300struct buffer_head *
1301__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1302{
1303 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1304
1305 if (bh == NULL) {
1306 /* __find_get_block_slow will mark the page accessed */
1307 bh = __find_get_block_slow(bdev, block);
1308 if (bh)
1309 bh_lru_install(bh);
1310 } else
1311 touch_buffer(bh);
1312
1313 return bh;
1314}
1315EXPORT_SYMBOL(__find_get_block);
1316
1317/*
1318 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1319 * which corresponds to the passed block_device, block and size. The
1320 * returned buffer has its reference count incremented.
1321 *
1322 * __getblk_gfp() will lock up the machine if grow_dev_page's
1323 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1324 */
1325struct buffer_head *
1326__getblk_gfp(struct block_device *bdev, sector_t block,
1327 unsigned size, gfp_t gfp)
1328{
1329 struct buffer_head *bh = __find_get_block(bdev, block, size);
1330
1331 might_sleep();
1332 if (bh == NULL)
1333 bh = __getblk_slow(bdev, block, size, gfp);
1334 return bh;
1335}
1336EXPORT_SYMBOL(__getblk_gfp);
1337
1338/*
1339 * Do async read-ahead on a buffer..
1340 */
1341void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1342{
1343 struct buffer_head *bh = __getblk(bdev, block, size);
1344 if (likely(bh)) {
1345 bh_readahead(bh, REQ_RAHEAD);
1346 brelse(bh);
1347 }
1348}
1349EXPORT_SYMBOL(__breadahead);
1350
1351/**
1352 * __bread_gfp() - reads a specified block and returns the bh
1353 * @bdev: the block_device to read from
1354 * @block: number of block
1355 * @size: size (in bytes) to read
1356 * @gfp: page allocation flag
1357 *
1358 * Reads a specified block, and returns buffer head that contains it.
1359 * The page cache can be allocated from non-movable area
1360 * not to prevent page migration if you set gfp to zero.
1361 * It returns NULL if the block was unreadable.
1362 */
1363struct buffer_head *
1364__bread_gfp(struct block_device *bdev, sector_t block,
1365 unsigned size, gfp_t gfp)
1366{
1367 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1368
1369 if (likely(bh) && !buffer_uptodate(bh))
1370 bh = __bread_slow(bh);
1371 return bh;
1372}
1373EXPORT_SYMBOL(__bread_gfp);
1374
1375static void __invalidate_bh_lrus(struct bh_lru *b)
1376{
1377 int i;
1378
1379 for (i = 0; i < BH_LRU_SIZE; i++) {
1380 brelse(b->bhs[i]);
1381 b->bhs[i] = NULL;
1382 }
1383}
1384/*
1385 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1386 * This doesn't race because it runs in each cpu either in irq
1387 * or with preempt disabled.
1388 */
1389static void invalidate_bh_lru(void *arg)
1390{
1391 struct bh_lru *b = &get_cpu_var(bh_lrus);
1392
1393 __invalidate_bh_lrus(b);
1394 put_cpu_var(bh_lrus);
1395}
1396
1397bool has_bh_in_lru(int cpu, void *dummy)
1398{
1399 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1400 int i;
1401
1402 for (i = 0; i < BH_LRU_SIZE; i++) {
1403 if (b->bhs[i])
1404 return true;
1405 }
1406
1407 return false;
1408}
1409
1410void invalidate_bh_lrus(void)
1411{
1412 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1413}
1414EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1415
1416/*
1417 * It's called from workqueue context so we need a bh_lru_lock to close
1418 * the race with preemption/irq.
1419 */
1420void invalidate_bh_lrus_cpu(void)
1421{
1422 struct bh_lru *b;
1423
1424 bh_lru_lock();
1425 b = this_cpu_ptr(&bh_lrus);
1426 __invalidate_bh_lrus(b);
1427 bh_lru_unlock();
1428}
1429
1430void set_bh_page(struct buffer_head *bh,
1431 struct page *page, unsigned long offset)
1432{
1433 bh->b_page = page;
1434 BUG_ON(offset >= PAGE_SIZE);
1435 if (PageHighMem(page))
1436 /*
1437 * This catches illegal uses and preserves the offset:
1438 */
1439 bh->b_data = (char *)(0 + offset);
1440 else
1441 bh->b_data = page_address(page) + offset;
1442}
1443EXPORT_SYMBOL(set_bh_page);
1444
1445/*
1446 * Called when truncating a buffer on a page completely.
1447 */
1448
1449/* Bits that are cleared during an invalidate */
1450#define BUFFER_FLAGS_DISCARD \
1451 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1452 1 << BH_Delay | 1 << BH_Unwritten)
1453
1454static void discard_buffer(struct buffer_head * bh)
1455{
1456 unsigned long b_state;
1457
1458 lock_buffer(bh);
1459 clear_buffer_dirty(bh);
1460 bh->b_bdev = NULL;
1461 b_state = READ_ONCE(bh->b_state);
1462 do {
1463 } while (!try_cmpxchg(&bh->b_state, &b_state,
1464 b_state & ~BUFFER_FLAGS_DISCARD));
1465 unlock_buffer(bh);
1466}
1467
1468/**
1469 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1470 * @folio: The folio which is affected.
1471 * @offset: start of the range to invalidate
1472 * @length: length of the range to invalidate
1473 *
1474 * block_invalidate_folio() is called when all or part of the folio has been
1475 * invalidated by a truncate operation.
1476 *
1477 * block_invalidate_folio() does not have to release all buffers, but it must
1478 * ensure that no dirty buffer is left outside @offset and that no I/O
1479 * is underway against any of the blocks which are outside the truncation
1480 * point. Because the caller is about to free (and possibly reuse) those
1481 * blocks on-disk.
1482 */
1483void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1484{
1485 struct buffer_head *head, *bh, *next;
1486 size_t curr_off = 0;
1487 size_t stop = length + offset;
1488
1489 BUG_ON(!folio_test_locked(folio));
1490
1491 /*
1492 * Check for overflow
1493 */
1494 BUG_ON(stop > folio_size(folio) || stop < length);
1495
1496 head = folio_buffers(folio);
1497 if (!head)
1498 return;
1499
1500 bh = head;
1501 do {
1502 size_t next_off = curr_off + bh->b_size;
1503 next = bh->b_this_page;
1504
1505 /*
1506 * Are we still fully in range ?
1507 */
1508 if (next_off > stop)
1509 goto out;
1510
1511 /*
1512 * is this block fully invalidated?
1513 */
1514 if (offset <= curr_off)
1515 discard_buffer(bh);
1516 curr_off = next_off;
1517 bh = next;
1518 } while (bh != head);
1519
1520 /*
1521 * We release buffers only if the entire folio is being invalidated.
1522 * The get_block cached value has been unconditionally invalidated,
1523 * so real IO is not possible anymore.
1524 */
1525 if (length == folio_size(folio))
1526 filemap_release_folio(folio, 0);
1527out:
1528 return;
1529}
1530EXPORT_SYMBOL(block_invalidate_folio);
1531
1532
1533/*
1534 * We attach and possibly dirty the buffers atomically wrt
1535 * block_dirty_folio() via private_lock. try_to_free_buffers
1536 * is already excluded via the page lock.
1537 */
1538void create_empty_buffers(struct page *page,
1539 unsigned long blocksize, unsigned long b_state)
1540{
1541 struct buffer_head *bh, *head, *tail;
1542
1543 head = alloc_page_buffers(page, blocksize, true);
1544 bh = head;
1545 do {
1546 bh->b_state |= b_state;
1547 tail = bh;
1548 bh = bh->b_this_page;
1549 } while (bh);
1550 tail->b_this_page = head;
1551
1552 spin_lock(&page->mapping->private_lock);
1553 if (PageUptodate(page) || PageDirty(page)) {
1554 bh = head;
1555 do {
1556 if (PageDirty(page))
1557 set_buffer_dirty(bh);
1558 if (PageUptodate(page))
1559 set_buffer_uptodate(bh);
1560 bh = bh->b_this_page;
1561 } while (bh != head);
1562 }
1563 attach_page_private(page, head);
1564 spin_unlock(&page->mapping->private_lock);
1565}
1566EXPORT_SYMBOL(create_empty_buffers);
1567
1568/**
1569 * clean_bdev_aliases: clean a range of buffers in block device
1570 * @bdev: Block device to clean buffers in
1571 * @block: Start of a range of blocks to clean
1572 * @len: Number of blocks to clean
1573 *
1574 * We are taking a range of blocks for data and we don't want writeback of any
1575 * buffer-cache aliases starting from return from this function and until the
1576 * moment when something will explicitly mark the buffer dirty (hopefully that
1577 * will not happen until we will free that block ;-) We don't even need to mark
1578 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1579 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1580 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1581 * would confuse anyone who might pick it with bread() afterwards...
1582 *
1583 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1584 * writeout I/O going on against recently-freed buffers. We don't wait on that
1585 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1586 * need to. That happens here.
1587 */
1588void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1589{
1590 struct inode *bd_inode = bdev->bd_inode;
1591 struct address_space *bd_mapping = bd_inode->i_mapping;
1592 struct folio_batch fbatch;
1593 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1594 pgoff_t end;
1595 int i, count;
1596 struct buffer_head *bh;
1597 struct buffer_head *head;
1598
1599 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1600 folio_batch_init(&fbatch);
1601 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1602 count = folio_batch_count(&fbatch);
1603 for (i = 0; i < count; i++) {
1604 struct folio *folio = fbatch.folios[i];
1605
1606 if (!folio_buffers(folio))
1607 continue;
1608 /*
1609 * We use folio lock instead of bd_mapping->private_lock
1610 * to pin buffers here since we can afford to sleep and
1611 * it scales better than a global spinlock lock.
1612 */
1613 folio_lock(folio);
1614 /* Recheck when the folio is locked which pins bhs */
1615 head = folio_buffers(folio);
1616 if (!head)
1617 goto unlock_page;
1618 bh = head;
1619 do {
1620 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1621 goto next;
1622 if (bh->b_blocknr >= block + len)
1623 break;
1624 clear_buffer_dirty(bh);
1625 wait_on_buffer(bh);
1626 clear_buffer_req(bh);
1627next:
1628 bh = bh->b_this_page;
1629 } while (bh != head);
1630unlock_page:
1631 folio_unlock(folio);
1632 }
1633 folio_batch_release(&fbatch);
1634 cond_resched();
1635 /* End of range already reached? */
1636 if (index > end || !index)
1637 break;
1638 }
1639}
1640EXPORT_SYMBOL(clean_bdev_aliases);
1641
1642/*
1643 * Size is a power-of-two in the range 512..PAGE_SIZE,
1644 * and the case we care about most is PAGE_SIZE.
1645 *
1646 * So this *could* possibly be written with those
1647 * constraints in mind (relevant mostly if some
1648 * architecture has a slow bit-scan instruction)
1649 */
1650static inline int block_size_bits(unsigned int blocksize)
1651{
1652 return ilog2(blocksize);
1653}
1654
1655static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1656{
1657 BUG_ON(!PageLocked(page));
1658
1659 if (!page_has_buffers(page))
1660 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1661 b_state);
1662 return page_buffers(page);
1663}
1664
1665/*
1666 * NOTE! All mapped/uptodate combinations are valid:
1667 *
1668 * Mapped Uptodate Meaning
1669 *
1670 * No No "unknown" - must do get_block()
1671 * No Yes "hole" - zero-filled
1672 * Yes No "allocated" - allocated on disk, not read in
1673 * Yes Yes "valid" - allocated and up-to-date in memory.
1674 *
1675 * "Dirty" is valid only with the last case (mapped+uptodate).
1676 */
1677
1678/*
1679 * While block_write_full_page is writing back the dirty buffers under
1680 * the page lock, whoever dirtied the buffers may decide to clean them
1681 * again at any time. We handle that by only looking at the buffer
1682 * state inside lock_buffer().
1683 *
1684 * If block_write_full_page() is called for regular writeback
1685 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1686 * locked buffer. This only can happen if someone has written the buffer
1687 * directly, with submit_bh(). At the address_space level PageWriteback
1688 * prevents this contention from occurring.
1689 *
1690 * If block_write_full_page() is called with wbc->sync_mode ==
1691 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1692 * causes the writes to be flagged as synchronous writes.
1693 */
1694int __block_write_full_page(struct inode *inode, struct page *page,
1695 get_block_t *get_block, struct writeback_control *wbc,
1696 bh_end_io_t *handler)
1697{
1698 int err;
1699 sector_t block;
1700 sector_t last_block;
1701 struct buffer_head *bh, *head;
1702 unsigned int blocksize, bbits;
1703 int nr_underway = 0;
1704 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1705
1706 head = create_page_buffers(page, inode,
1707 (1 << BH_Dirty)|(1 << BH_Uptodate));
1708
1709 /*
1710 * Be very careful. We have no exclusion from block_dirty_folio
1711 * here, and the (potentially unmapped) buffers may become dirty at
1712 * any time. If a buffer becomes dirty here after we've inspected it
1713 * then we just miss that fact, and the page stays dirty.
1714 *
1715 * Buffers outside i_size may be dirtied by block_dirty_folio;
1716 * handle that here by just cleaning them.
1717 */
1718
1719 bh = head;
1720 blocksize = bh->b_size;
1721 bbits = block_size_bits(blocksize);
1722
1723 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1724 last_block = (i_size_read(inode) - 1) >> bbits;
1725
1726 /*
1727 * Get all the dirty buffers mapped to disk addresses and
1728 * handle any aliases from the underlying blockdev's mapping.
1729 */
1730 do {
1731 if (block > last_block) {
1732 /*
1733 * mapped buffers outside i_size will occur, because
1734 * this page can be outside i_size when there is a
1735 * truncate in progress.
1736 */
1737 /*
1738 * The buffer was zeroed by block_write_full_page()
1739 */
1740 clear_buffer_dirty(bh);
1741 set_buffer_uptodate(bh);
1742 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1743 buffer_dirty(bh)) {
1744 WARN_ON(bh->b_size != blocksize);
1745 err = get_block(inode, block, bh, 1);
1746 if (err)
1747 goto recover;
1748 clear_buffer_delay(bh);
1749 if (buffer_new(bh)) {
1750 /* blockdev mappings never come here */
1751 clear_buffer_new(bh);
1752 clean_bdev_bh_alias(bh);
1753 }
1754 }
1755 bh = bh->b_this_page;
1756 block++;
1757 } while (bh != head);
1758
1759 do {
1760 if (!buffer_mapped(bh))
1761 continue;
1762 /*
1763 * If it's a fully non-blocking write attempt and we cannot
1764 * lock the buffer then redirty the page. Note that this can
1765 * potentially cause a busy-wait loop from writeback threads
1766 * and kswapd activity, but those code paths have their own
1767 * higher-level throttling.
1768 */
1769 if (wbc->sync_mode != WB_SYNC_NONE) {
1770 lock_buffer(bh);
1771 } else if (!trylock_buffer(bh)) {
1772 redirty_page_for_writepage(wbc, page);
1773 continue;
1774 }
1775 if (test_clear_buffer_dirty(bh)) {
1776 mark_buffer_async_write_endio(bh, handler);
1777 } else {
1778 unlock_buffer(bh);
1779 }
1780 } while ((bh = bh->b_this_page) != head);
1781
1782 /*
1783 * The page and its buffers are protected by PageWriteback(), so we can
1784 * drop the bh refcounts early.
1785 */
1786 BUG_ON(PageWriteback(page));
1787 set_page_writeback(page);
1788
1789 do {
1790 struct buffer_head *next = bh->b_this_page;
1791 if (buffer_async_write(bh)) {
1792 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1793 nr_underway++;
1794 }
1795 bh = next;
1796 } while (bh != head);
1797 unlock_page(page);
1798
1799 err = 0;
1800done:
1801 if (nr_underway == 0) {
1802 /*
1803 * The page was marked dirty, but the buffers were
1804 * clean. Someone wrote them back by hand with
1805 * write_dirty_buffer/submit_bh. A rare case.
1806 */
1807 end_page_writeback(page);
1808
1809 /*
1810 * The page and buffer_heads can be released at any time from
1811 * here on.
1812 */
1813 }
1814 return err;
1815
1816recover:
1817 /*
1818 * ENOSPC, or some other error. We may already have added some
1819 * blocks to the file, so we need to write these out to avoid
1820 * exposing stale data.
1821 * The page is currently locked and not marked for writeback
1822 */
1823 bh = head;
1824 /* Recovery: lock and submit the mapped buffers */
1825 do {
1826 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1827 !buffer_delay(bh)) {
1828 lock_buffer(bh);
1829 mark_buffer_async_write_endio(bh, handler);
1830 } else {
1831 /*
1832 * The buffer may have been set dirty during
1833 * attachment to a dirty page.
1834 */
1835 clear_buffer_dirty(bh);
1836 }
1837 } while ((bh = bh->b_this_page) != head);
1838 SetPageError(page);
1839 BUG_ON(PageWriteback(page));
1840 mapping_set_error(page->mapping, err);
1841 set_page_writeback(page);
1842 do {
1843 struct buffer_head *next = bh->b_this_page;
1844 if (buffer_async_write(bh)) {
1845 clear_buffer_dirty(bh);
1846 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1847 nr_underway++;
1848 }
1849 bh = next;
1850 } while (bh != head);
1851 unlock_page(page);
1852 goto done;
1853}
1854EXPORT_SYMBOL(__block_write_full_page);
1855
1856/*
1857 * If a page has any new buffers, zero them out here, and mark them uptodate
1858 * and dirty so they'll be written out (in order to prevent uninitialised
1859 * block data from leaking). And clear the new bit.
1860 */
1861void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1862{
1863 unsigned int block_start, block_end;
1864 struct buffer_head *head, *bh;
1865
1866 BUG_ON(!PageLocked(page));
1867 if (!page_has_buffers(page))
1868 return;
1869
1870 bh = head = page_buffers(page);
1871 block_start = 0;
1872 do {
1873 block_end = block_start + bh->b_size;
1874
1875 if (buffer_new(bh)) {
1876 if (block_end > from && block_start < to) {
1877 if (!PageUptodate(page)) {
1878 unsigned start, size;
1879
1880 start = max(from, block_start);
1881 size = min(to, block_end) - start;
1882
1883 zero_user(page, start, size);
1884 set_buffer_uptodate(bh);
1885 }
1886
1887 clear_buffer_new(bh);
1888 mark_buffer_dirty(bh);
1889 }
1890 }
1891
1892 block_start = block_end;
1893 bh = bh->b_this_page;
1894 } while (bh != head);
1895}
1896EXPORT_SYMBOL(page_zero_new_buffers);
1897
1898static void
1899iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1900 const struct iomap *iomap)
1901{
1902 loff_t offset = block << inode->i_blkbits;
1903
1904 bh->b_bdev = iomap->bdev;
1905
1906 /*
1907 * Block points to offset in file we need to map, iomap contains
1908 * the offset at which the map starts. If the map ends before the
1909 * current block, then do not map the buffer and let the caller
1910 * handle it.
1911 */
1912 BUG_ON(offset >= iomap->offset + iomap->length);
1913
1914 switch (iomap->type) {
1915 case IOMAP_HOLE:
1916 /*
1917 * If the buffer is not up to date or beyond the current EOF,
1918 * we need to mark it as new to ensure sub-block zeroing is
1919 * executed if necessary.
1920 */
1921 if (!buffer_uptodate(bh) ||
1922 (offset >= i_size_read(inode)))
1923 set_buffer_new(bh);
1924 break;
1925 case IOMAP_DELALLOC:
1926 if (!buffer_uptodate(bh) ||
1927 (offset >= i_size_read(inode)))
1928 set_buffer_new(bh);
1929 set_buffer_uptodate(bh);
1930 set_buffer_mapped(bh);
1931 set_buffer_delay(bh);
1932 break;
1933 case IOMAP_UNWRITTEN:
1934 /*
1935 * For unwritten regions, we always need to ensure that regions
1936 * in the block we are not writing to are zeroed. Mark the
1937 * buffer as new to ensure this.
1938 */
1939 set_buffer_new(bh);
1940 set_buffer_unwritten(bh);
1941 fallthrough;
1942 case IOMAP_MAPPED:
1943 if ((iomap->flags & IOMAP_F_NEW) ||
1944 offset >= i_size_read(inode))
1945 set_buffer_new(bh);
1946 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1947 inode->i_blkbits;
1948 set_buffer_mapped(bh);
1949 break;
1950 }
1951}
1952
1953int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
1954 get_block_t *get_block, const struct iomap *iomap)
1955{
1956 unsigned from = pos & (PAGE_SIZE - 1);
1957 unsigned to = from + len;
1958 struct inode *inode = folio->mapping->host;
1959 unsigned block_start, block_end;
1960 sector_t block;
1961 int err = 0;
1962 unsigned blocksize, bbits;
1963 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1964
1965 BUG_ON(!folio_test_locked(folio));
1966 BUG_ON(from > PAGE_SIZE);
1967 BUG_ON(to > PAGE_SIZE);
1968 BUG_ON(from > to);
1969
1970 head = create_page_buffers(&folio->page, inode, 0);
1971 blocksize = head->b_size;
1972 bbits = block_size_bits(blocksize);
1973
1974 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
1975
1976 for(bh = head, block_start = 0; bh != head || !block_start;
1977 block++, block_start=block_end, bh = bh->b_this_page) {
1978 block_end = block_start + blocksize;
1979 if (block_end <= from || block_start >= to) {
1980 if (folio_test_uptodate(folio)) {
1981 if (!buffer_uptodate(bh))
1982 set_buffer_uptodate(bh);
1983 }
1984 continue;
1985 }
1986 if (buffer_new(bh))
1987 clear_buffer_new(bh);
1988 if (!buffer_mapped(bh)) {
1989 WARN_ON(bh->b_size != blocksize);
1990 if (get_block) {
1991 err = get_block(inode, block, bh, 1);
1992 if (err)
1993 break;
1994 } else {
1995 iomap_to_bh(inode, block, bh, iomap);
1996 }
1997
1998 if (buffer_new(bh)) {
1999 clean_bdev_bh_alias(bh);
2000 if (folio_test_uptodate(folio)) {
2001 clear_buffer_new(bh);
2002 set_buffer_uptodate(bh);
2003 mark_buffer_dirty(bh);
2004 continue;
2005 }
2006 if (block_end > to || block_start < from)
2007 folio_zero_segments(folio,
2008 to, block_end,
2009 block_start, from);
2010 continue;
2011 }
2012 }
2013 if (folio_test_uptodate(folio)) {
2014 if (!buffer_uptodate(bh))
2015 set_buffer_uptodate(bh);
2016 continue;
2017 }
2018 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2019 !buffer_unwritten(bh) &&
2020 (block_start < from || block_end > to)) {
2021 bh_read_nowait(bh, 0);
2022 *wait_bh++=bh;
2023 }
2024 }
2025 /*
2026 * If we issued read requests - let them complete.
2027 */
2028 while(wait_bh > wait) {
2029 wait_on_buffer(*--wait_bh);
2030 if (!buffer_uptodate(*wait_bh))
2031 err = -EIO;
2032 }
2033 if (unlikely(err))
2034 page_zero_new_buffers(&folio->page, from, to);
2035 return err;
2036}
2037
2038int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2039 get_block_t *get_block)
2040{
2041 return __block_write_begin_int(page_folio(page), pos, len, get_block,
2042 NULL);
2043}
2044EXPORT_SYMBOL(__block_write_begin);
2045
2046static int __block_commit_write(struct inode *inode, struct page *page,
2047 unsigned from, unsigned to)
2048{
2049 unsigned block_start, block_end;
2050 int partial = 0;
2051 unsigned blocksize;
2052 struct buffer_head *bh, *head;
2053
2054 bh = head = page_buffers(page);
2055 blocksize = bh->b_size;
2056
2057 block_start = 0;
2058 do {
2059 block_end = block_start + blocksize;
2060 if (block_end <= from || block_start >= to) {
2061 if (!buffer_uptodate(bh))
2062 partial = 1;
2063 } else {
2064 set_buffer_uptodate(bh);
2065 mark_buffer_dirty(bh);
2066 }
2067 if (buffer_new(bh))
2068 clear_buffer_new(bh);
2069
2070 block_start = block_end;
2071 bh = bh->b_this_page;
2072 } while (bh != head);
2073
2074 /*
2075 * If this is a partial write which happened to make all buffers
2076 * uptodate then we can optimize away a bogus read_folio() for
2077 * the next read(). Here we 'discover' whether the page went
2078 * uptodate as a result of this (potentially partial) write.
2079 */
2080 if (!partial)
2081 SetPageUptodate(page);
2082 return 0;
2083}
2084
2085/*
2086 * block_write_begin takes care of the basic task of block allocation and
2087 * bringing partial write blocks uptodate first.
2088 *
2089 * The filesystem needs to handle block truncation upon failure.
2090 */
2091int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2092 struct page **pagep, get_block_t *get_block)
2093{
2094 pgoff_t index = pos >> PAGE_SHIFT;
2095 struct page *page;
2096 int status;
2097
2098 page = grab_cache_page_write_begin(mapping, index);
2099 if (!page)
2100 return -ENOMEM;
2101
2102 status = __block_write_begin(page, pos, len, get_block);
2103 if (unlikely(status)) {
2104 unlock_page(page);
2105 put_page(page);
2106 page = NULL;
2107 }
2108
2109 *pagep = page;
2110 return status;
2111}
2112EXPORT_SYMBOL(block_write_begin);
2113
2114int block_write_end(struct file *file, struct address_space *mapping,
2115 loff_t pos, unsigned len, unsigned copied,
2116 struct page *page, void *fsdata)
2117{
2118 struct inode *inode = mapping->host;
2119 unsigned start;
2120
2121 start = pos & (PAGE_SIZE - 1);
2122
2123 if (unlikely(copied < len)) {
2124 /*
2125 * The buffers that were written will now be uptodate, so
2126 * we don't have to worry about a read_folio reading them
2127 * and overwriting a partial write. However if we have
2128 * encountered a short write and only partially written
2129 * into a buffer, it will not be marked uptodate, so a
2130 * read_folio might come in and destroy our partial write.
2131 *
2132 * Do the simplest thing, and just treat any short write to a
2133 * non uptodate page as a zero-length write, and force the
2134 * caller to redo the whole thing.
2135 */
2136 if (!PageUptodate(page))
2137 copied = 0;
2138
2139 page_zero_new_buffers(page, start+copied, start+len);
2140 }
2141 flush_dcache_page(page);
2142
2143 /* This could be a short (even 0-length) commit */
2144 __block_commit_write(inode, page, start, start+copied);
2145
2146 return copied;
2147}
2148EXPORT_SYMBOL(block_write_end);
2149
2150int generic_write_end(struct file *file, struct address_space *mapping,
2151 loff_t pos, unsigned len, unsigned copied,
2152 struct page *page, void *fsdata)
2153{
2154 struct inode *inode = mapping->host;
2155 loff_t old_size = inode->i_size;
2156 bool i_size_changed = false;
2157
2158 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2159
2160 /*
2161 * No need to use i_size_read() here, the i_size cannot change under us
2162 * because we hold i_rwsem.
2163 *
2164 * But it's important to update i_size while still holding page lock:
2165 * page writeout could otherwise come in and zero beyond i_size.
2166 */
2167 if (pos + copied > inode->i_size) {
2168 i_size_write(inode, pos + copied);
2169 i_size_changed = true;
2170 }
2171
2172 unlock_page(page);
2173 put_page(page);
2174
2175 if (old_size < pos)
2176 pagecache_isize_extended(inode, old_size, pos);
2177 /*
2178 * Don't mark the inode dirty under page lock. First, it unnecessarily
2179 * makes the holding time of page lock longer. Second, it forces lock
2180 * ordering of page lock and transaction start for journaling
2181 * filesystems.
2182 */
2183 if (i_size_changed)
2184 mark_inode_dirty(inode);
2185 return copied;
2186}
2187EXPORT_SYMBOL(generic_write_end);
2188
2189/*
2190 * block_is_partially_uptodate checks whether buffers within a folio are
2191 * uptodate or not.
2192 *
2193 * Returns true if all buffers which correspond to the specified part
2194 * of the folio are uptodate.
2195 */
2196bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2197{
2198 unsigned block_start, block_end, blocksize;
2199 unsigned to;
2200 struct buffer_head *bh, *head;
2201 bool ret = true;
2202
2203 head = folio_buffers(folio);
2204 if (!head)
2205 return false;
2206 blocksize = head->b_size;
2207 to = min_t(unsigned, folio_size(folio) - from, count);
2208 to = from + to;
2209 if (from < blocksize && to > folio_size(folio) - blocksize)
2210 return false;
2211
2212 bh = head;
2213 block_start = 0;
2214 do {
2215 block_end = block_start + blocksize;
2216 if (block_end > from && block_start < to) {
2217 if (!buffer_uptodate(bh)) {
2218 ret = false;
2219 break;
2220 }
2221 if (block_end >= to)
2222 break;
2223 }
2224 block_start = block_end;
2225 bh = bh->b_this_page;
2226 } while (bh != head);
2227
2228 return ret;
2229}
2230EXPORT_SYMBOL(block_is_partially_uptodate);
2231
2232/*
2233 * Generic "read_folio" function for block devices that have the normal
2234 * get_block functionality. This is most of the block device filesystems.
2235 * Reads the folio asynchronously --- the unlock_buffer() and
2236 * set/clear_buffer_uptodate() functions propagate buffer state into the
2237 * folio once IO has completed.
2238 */
2239int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2240{
2241 struct inode *inode = folio->mapping->host;
2242 sector_t iblock, lblock;
2243 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2244 unsigned int blocksize, bbits;
2245 int nr, i;
2246 int fully_mapped = 1;
2247 bool page_error = false;
2248
2249 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2250
2251 head = create_page_buffers(&folio->page, inode, 0);
2252 blocksize = head->b_size;
2253 bbits = block_size_bits(blocksize);
2254
2255 iblock = (sector_t)folio->index << (PAGE_SHIFT - bbits);
2256 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2257 bh = head;
2258 nr = 0;
2259 i = 0;
2260
2261 do {
2262 if (buffer_uptodate(bh))
2263 continue;
2264
2265 if (!buffer_mapped(bh)) {
2266 int err = 0;
2267
2268 fully_mapped = 0;
2269 if (iblock < lblock) {
2270 WARN_ON(bh->b_size != blocksize);
2271 err = get_block(inode, iblock, bh, 0);
2272 if (err) {
2273 folio_set_error(folio);
2274 page_error = true;
2275 }
2276 }
2277 if (!buffer_mapped(bh)) {
2278 folio_zero_range(folio, i * blocksize,
2279 blocksize);
2280 if (!err)
2281 set_buffer_uptodate(bh);
2282 continue;
2283 }
2284 /*
2285 * get_block() might have updated the buffer
2286 * synchronously
2287 */
2288 if (buffer_uptodate(bh))
2289 continue;
2290 }
2291 arr[nr++] = bh;
2292 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2293
2294 if (fully_mapped)
2295 folio_set_mappedtodisk(folio);
2296
2297 if (!nr) {
2298 /*
2299 * All buffers are uptodate - we can set the folio uptodate
2300 * as well. But not if get_block() returned an error.
2301 */
2302 if (!page_error)
2303 folio_mark_uptodate(folio);
2304 folio_unlock(folio);
2305 return 0;
2306 }
2307
2308 /* Stage two: lock the buffers */
2309 for (i = 0; i < nr; i++) {
2310 bh = arr[i];
2311 lock_buffer(bh);
2312 mark_buffer_async_read(bh);
2313 }
2314
2315 /*
2316 * Stage 3: start the IO. Check for uptodateness
2317 * inside the buffer lock in case another process reading
2318 * the underlying blockdev brought it uptodate (the sct fix).
2319 */
2320 for (i = 0; i < nr; i++) {
2321 bh = arr[i];
2322 if (buffer_uptodate(bh))
2323 end_buffer_async_read(bh, 1);
2324 else
2325 submit_bh(REQ_OP_READ, bh);
2326 }
2327 return 0;
2328}
2329EXPORT_SYMBOL(block_read_full_folio);
2330
2331/* utility function for filesystems that need to do work on expanding
2332 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2333 * deal with the hole.
2334 */
2335int generic_cont_expand_simple(struct inode *inode, loff_t size)
2336{
2337 struct address_space *mapping = inode->i_mapping;
2338 const struct address_space_operations *aops = mapping->a_ops;
2339 struct page *page;
2340 void *fsdata = NULL;
2341 int err;
2342
2343 err = inode_newsize_ok(inode, size);
2344 if (err)
2345 goto out;
2346
2347 err = aops->write_begin(NULL, mapping, size, 0, &page, &fsdata);
2348 if (err)
2349 goto out;
2350
2351 err = aops->write_end(NULL, mapping, size, 0, 0, page, fsdata);
2352 BUG_ON(err > 0);
2353
2354out:
2355 return err;
2356}
2357EXPORT_SYMBOL(generic_cont_expand_simple);
2358
2359static int cont_expand_zero(struct file *file, struct address_space *mapping,
2360 loff_t pos, loff_t *bytes)
2361{
2362 struct inode *inode = mapping->host;
2363 const struct address_space_operations *aops = mapping->a_ops;
2364 unsigned int blocksize = i_blocksize(inode);
2365 struct page *page;
2366 void *fsdata = NULL;
2367 pgoff_t index, curidx;
2368 loff_t curpos;
2369 unsigned zerofrom, offset, len;
2370 int err = 0;
2371
2372 index = pos >> PAGE_SHIFT;
2373 offset = pos & ~PAGE_MASK;
2374
2375 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2376 zerofrom = curpos & ~PAGE_MASK;
2377 if (zerofrom & (blocksize-1)) {
2378 *bytes |= (blocksize-1);
2379 (*bytes)++;
2380 }
2381 len = PAGE_SIZE - zerofrom;
2382
2383 err = aops->write_begin(file, mapping, curpos, len,
2384 &page, &fsdata);
2385 if (err)
2386 goto out;
2387 zero_user(page, zerofrom, len);
2388 err = aops->write_end(file, mapping, curpos, len, len,
2389 page, fsdata);
2390 if (err < 0)
2391 goto out;
2392 BUG_ON(err != len);
2393 err = 0;
2394
2395 balance_dirty_pages_ratelimited(mapping);
2396
2397 if (fatal_signal_pending(current)) {
2398 err = -EINTR;
2399 goto out;
2400 }
2401 }
2402
2403 /* page covers the boundary, find the boundary offset */
2404 if (index == curidx) {
2405 zerofrom = curpos & ~PAGE_MASK;
2406 /* if we will expand the thing last block will be filled */
2407 if (offset <= zerofrom) {
2408 goto out;
2409 }
2410 if (zerofrom & (blocksize-1)) {
2411 *bytes |= (blocksize-1);
2412 (*bytes)++;
2413 }
2414 len = offset - zerofrom;
2415
2416 err = aops->write_begin(file, mapping, curpos, len,
2417 &page, &fsdata);
2418 if (err)
2419 goto out;
2420 zero_user(page, zerofrom, len);
2421 err = aops->write_end(file, mapping, curpos, len, len,
2422 page, fsdata);
2423 if (err < 0)
2424 goto out;
2425 BUG_ON(err != len);
2426 err = 0;
2427 }
2428out:
2429 return err;
2430}
2431
2432/*
2433 * For moronic filesystems that do not allow holes in file.
2434 * We may have to extend the file.
2435 */
2436int cont_write_begin(struct file *file, struct address_space *mapping,
2437 loff_t pos, unsigned len,
2438 struct page **pagep, void **fsdata,
2439 get_block_t *get_block, loff_t *bytes)
2440{
2441 struct inode *inode = mapping->host;
2442 unsigned int blocksize = i_blocksize(inode);
2443 unsigned int zerofrom;
2444 int err;
2445
2446 err = cont_expand_zero(file, mapping, pos, bytes);
2447 if (err)
2448 return err;
2449
2450 zerofrom = *bytes & ~PAGE_MASK;
2451 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2452 *bytes |= (blocksize-1);
2453 (*bytes)++;
2454 }
2455
2456 return block_write_begin(mapping, pos, len, pagep, get_block);
2457}
2458EXPORT_SYMBOL(cont_write_begin);
2459
2460int block_commit_write(struct page *page, unsigned from, unsigned to)
2461{
2462 struct inode *inode = page->mapping->host;
2463 __block_commit_write(inode,page,from,to);
2464 return 0;
2465}
2466EXPORT_SYMBOL(block_commit_write);
2467
2468/*
2469 * block_page_mkwrite() is not allowed to change the file size as it gets
2470 * called from a page fault handler when a page is first dirtied. Hence we must
2471 * be careful to check for EOF conditions here. We set the page up correctly
2472 * for a written page which means we get ENOSPC checking when writing into
2473 * holes and correct delalloc and unwritten extent mapping on filesystems that
2474 * support these features.
2475 *
2476 * We are not allowed to take the i_mutex here so we have to play games to
2477 * protect against truncate races as the page could now be beyond EOF. Because
2478 * truncate writes the inode size before removing pages, once we have the
2479 * page lock we can determine safely if the page is beyond EOF. If it is not
2480 * beyond EOF, then the page is guaranteed safe against truncation until we
2481 * unlock the page.
2482 *
2483 * Direct callers of this function should protect against filesystem freezing
2484 * using sb_start_pagefault() - sb_end_pagefault() functions.
2485 */
2486int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2487 get_block_t get_block)
2488{
2489 struct page *page = vmf->page;
2490 struct inode *inode = file_inode(vma->vm_file);
2491 unsigned long end;
2492 loff_t size;
2493 int ret;
2494
2495 lock_page(page);
2496 size = i_size_read(inode);
2497 if ((page->mapping != inode->i_mapping) ||
2498 (page_offset(page) > size)) {
2499 /* We overload EFAULT to mean page got truncated */
2500 ret = -EFAULT;
2501 goto out_unlock;
2502 }
2503
2504 /* page is wholly or partially inside EOF */
2505 if (((page->index + 1) << PAGE_SHIFT) > size)
2506 end = size & ~PAGE_MASK;
2507 else
2508 end = PAGE_SIZE;
2509
2510 ret = __block_write_begin(page, 0, end, get_block);
2511 if (!ret)
2512 ret = block_commit_write(page, 0, end);
2513
2514 if (unlikely(ret < 0))
2515 goto out_unlock;
2516 set_page_dirty(page);
2517 wait_for_stable_page(page);
2518 return 0;
2519out_unlock:
2520 unlock_page(page);
2521 return ret;
2522}
2523EXPORT_SYMBOL(block_page_mkwrite);
2524
2525int block_truncate_page(struct address_space *mapping,
2526 loff_t from, get_block_t *get_block)
2527{
2528 pgoff_t index = from >> PAGE_SHIFT;
2529 unsigned offset = from & (PAGE_SIZE-1);
2530 unsigned blocksize;
2531 sector_t iblock;
2532 unsigned length, pos;
2533 struct inode *inode = mapping->host;
2534 struct page *page;
2535 struct buffer_head *bh;
2536 int err;
2537
2538 blocksize = i_blocksize(inode);
2539 length = offset & (blocksize - 1);
2540
2541 /* Block boundary? Nothing to do */
2542 if (!length)
2543 return 0;
2544
2545 length = blocksize - length;
2546 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2547
2548 page = grab_cache_page(mapping, index);
2549 err = -ENOMEM;
2550 if (!page)
2551 goto out;
2552
2553 if (!page_has_buffers(page))
2554 create_empty_buffers(page, blocksize, 0);
2555
2556 /* Find the buffer that contains "offset" */
2557 bh = page_buffers(page);
2558 pos = blocksize;
2559 while (offset >= pos) {
2560 bh = bh->b_this_page;
2561 iblock++;
2562 pos += blocksize;
2563 }
2564
2565 err = 0;
2566 if (!buffer_mapped(bh)) {
2567 WARN_ON(bh->b_size != blocksize);
2568 err = get_block(inode, iblock, bh, 0);
2569 if (err)
2570 goto unlock;
2571 /* unmapped? It's a hole - nothing to do */
2572 if (!buffer_mapped(bh))
2573 goto unlock;
2574 }
2575
2576 /* Ok, it's mapped. Make sure it's up-to-date */
2577 if (PageUptodate(page))
2578 set_buffer_uptodate(bh);
2579
2580 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2581 err = bh_read(bh, 0);
2582 /* Uhhuh. Read error. Complain and punt. */
2583 if (err < 0)
2584 goto unlock;
2585 }
2586
2587 zero_user(page, offset, length);
2588 mark_buffer_dirty(bh);
2589 err = 0;
2590
2591unlock:
2592 unlock_page(page);
2593 put_page(page);
2594out:
2595 return err;
2596}
2597EXPORT_SYMBOL(block_truncate_page);
2598
2599/*
2600 * The generic ->writepage function for buffer-backed address_spaces
2601 */
2602int block_write_full_page(struct page *page, get_block_t *get_block,
2603 struct writeback_control *wbc)
2604{
2605 struct inode * const inode = page->mapping->host;
2606 loff_t i_size = i_size_read(inode);
2607 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2608 unsigned offset;
2609
2610 /* Is the page fully inside i_size? */
2611 if (page->index < end_index)
2612 return __block_write_full_page(inode, page, get_block, wbc,
2613 end_buffer_async_write);
2614
2615 /* Is the page fully outside i_size? (truncate in progress) */
2616 offset = i_size & (PAGE_SIZE-1);
2617 if (page->index >= end_index+1 || !offset) {
2618 unlock_page(page);
2619 return 0; /* don't care */
2620 }
2621
2622 /*
2623 * The page straddles i_size. It must be zeroed out on each and every
2624 * writepage invocation because it may be mmapped. "A file is mapped
2625 * in multiples of the page size. For a file that is not a multiple of
2626 * the page size, the remaining memory is zeroed when mapped, and
2627 * writes to that region are not written out to the file."
2628 */
2629 zero_user_segment(page, offset, PAGE_SIZE);
2630 return __block_write_full_page(inode, page, get_block, wbc,
2631 end_buffer_async_write);
2632}
2633EXPORT_SYMBOL(block_write_full_page);
2634
2635sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2636 get_block_t *get_block)
2637{
2638 struct inode *inode = mapping->host;
2639 struct buffer_head tmp = {
2640 .b_size = i_blocksize(inode),
2641 };
2642
2643 get_block(inode, block, &tmp, 0);
2644 return tmp.b_blocknr;
2645}
2646EXPORT_SYMBOL(generic_block_bmap);
2647
2648static void end_bio_bh_io_sync(struct bio *bio)
2649{
2650 struct buffer_head *bh = bio->bi_private;
2651
2652 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2653 set_bit(BH_Quiet, &bh->b_state);
2654
2655 bh->b_end_io(bh, !bio->bi_status);
2656 bio_put(bio);
2657}
2658
2659static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2660 struct writeback_control *wbc)
2661{
2662 const enum req_op op = opf & REQ_OP_MASK;
2663 struct bio *bio;
2664
2665 BUG_ON(!buffer_locked(bh));
2666 BUG_ON(!buffer_mapped(bh));
2667 BUG_ON(!bh->b_end_io);
2668 BUG_ON(buffer_delay(bh));
2669 BUG_ON(buffer_unwritten(bh));
2670
2671 /*
2672 * Only clear out a write error when rewriting
2673 */
2674 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2675 clear_buffer_write_io_error(bh);
2676
2677 if (buffer_meta(bh))
2678 opf |= REQ_META;
2679 if (buffer_prio(bh))
2680 opf |= REQ_PRIO;
2681
2682 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2683
2684 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2685
2686 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2687
2688 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
2689 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
2690
2691 bio->bi_end_io = end_bio_bh_io_sync;
2692 bio->bi_private = bh;
2693
2694 /* Take care of bh's that straddle the end of the device */
2695 guard_bio_eod(bio);
2696
2697 if (wbc) {
2698 wbc_init_bio(wbc, bio);
2699 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
2700 }
2701
2702 submit_bio(bio);
2703}
2704
2705void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2706{
2707 submit_bh_wbc(opf, bh, NULL);
2708}
2709EXPORT_SYMBOL(submit_bh);
2710
2711void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2712{
2713 lock_buffer(bh);
2714 if (!test_clear_buffer_dirty(bh)) {
2715 unlock_buffer(bh);
2716 return;
2717 }
2718 bh->b_end_io = end_buffer_write_sync;
2719 get_bh(bh);
2720 submit_bh(REQ_OP_WRITE | op_flags, bh);
2721}
2722EXPORT_SYMBOL(write_dirty_buffer);
2723
2724/*
2725 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2726 * and then start new I/O and then wait upon it. The caller must have a ref on
2727 * the buffer_head.
2728 */
2729int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2730{
2731 WARN_ON(atomic_read(&bh->b_count) < 1);
2732 lock_buffer(bh);
2733 if (test_clear_buffer_dirty(bh)) {
2734 /*
2735 * The bh should be mapped, but it might not be if the
2736 * device was hot-removed. Not much we can do but fail the I/O.
2737 */
2738 if (!buffer_mapped(bh)) {
2739 unlock_buffer(bh);
2740 return -EIO;
2741 }
2742
2743 get_bh(bh);
2744 bh->b_end_io = end_buffer_write_sync;
2745 submit_bh(REQ_OP_WRITE | op_flags, bh);
2746 wait_on_buffer(bh);
2747 if (!buffer_uptodate(bh))
2748 return -EIO;
2749 } else {
2750 unlock_buffer(bh);
2751 }
2752 return 0;
2753}
2754EXPORT_SYMBOL(__sync_dirty_buffer);
2755
2756int sync_dirty_buffer(struct buffer_head *bh)
2757{
2758 return __sync_dirty_buffer(bh, REQ_SYNC);
2759}
2760EXPORT_SYMBOL(sync_dirty_buffer);
2761
2762/*
2763 * try_to_free_buffers() checks if all the buffers on this particular folio
2764 * are unused, and releases them if so.
2765 *
2766 * Exclusion against try_to_free_buffers may be obtained by either
2767 * locking the folio or by holding its mapping's private_lock.
2768 *
2769 * If the folio is dirty but all the buffers are clean then we need to
2770 * be sure to mark the folio clean as well. This is because the folio
2771 * may be against a block device, and a later reattachment of buffers
2772 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2773 * filesystem data on the same device.
2774 *
2775 * The same applies to regular filesystem folios: if all the buffers are
2776 * clean then we set the folio clean and proceed. To do that, we require
2777 * total exclusion from block_dirty_folio(). That is obtained with
2778 * private_lock.
2779 *
2780 * try_to_free_buffers() is non-blocking.
2781 */
2782static inline int buffer_busy(struct buffer_head *bh)
2783{
2784 return atomic_read(&bh->b_count) |
2785 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2786}
2787
2788static bool
2789drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2790{
2791 struct buffer_head *head = folio_buffers(folio);
2792 struct buffer_head *bh;
2793
2794 bh = head;
2795 do {
2796 if (buffer_busy(bh))
2797 goto failed;
2798 bh = bh->b_this_page;
2799 } while (bh != head);
2800
2801 do {
2802 struct buffer_head *next = bh->b_this_page;
2803
2804 if (bh->b_assoc_map)
2805 __remove_assoc_queue(bh);
2806 bh = next;
2807 } while (bh != head);
2808 *buffers_to_free = head;
2809 folio_detach_private(folio);
2810 return true;
2811failed:
2812 return false;
2813}
2814
2815bool try_to_free_buffers(struct folio *folio)
2816{
2817 struct address_space * const mapping = folio->mapping;
2818 struct buffer_head *buffers_to_free = NULL;
2819 bool ret = 0;
2820
2821 BUG_ON(!folio_test_locked(folio));
2822 if (folio_test_writeback(folio))
2823 return false;
2824
2825 if (mapping == NULL) { /* can this still happen? */
2826 ret = drop_buffers(folio, &buffers_to_free);
2827 goto out;
2828 }
2829
2830 spin_lock(&mapping->private_lock);
2831 ret = drop_buffers(folio, &buffers_to_free);
2832
2833 /*
2834 * If the filesystem writes its buffers by hand (eg ext3)
2835 * then we can have clean buffers against a dirty folio. We
2836 * clean the folio here; otherwise the VM will never notice
2837 * that the filesystem did any IO at all.
2838 *
2839 * Also, during truncate, discard_buffer will have marked all
2840 * the folio's buffers clean. We discover that here and clean
2841 * the folio also.
2842 *
2843 * private_lock must be held over this entire operation in order
2844 * to synchronise against block_dirty_folio and prevent the
2845 * dirty bit from being lost.
2846 */
2847 if (ret)
2848 folio_cancel_dirty(folio);
2849 spin_unlock(&mapping->private_lock);
2850out:
2851 if (buffers_to_free) {
2852 struct buffer_head *bh = buffers_to_free;
2853
2854 do {
2855 struct buffer_head *next = bh->b_this_page;
2856 free_buffer_head(bh);
2857 bh = next;
2858 } while (bh != buffers_to_free);
2859 }
2860 return ret;
2861}
2862EXPORT_SYMBOL(try_to_free_buffers);
2863
2864/*
2865 * Buffer-head allocation
2866 */
2867static struct kmem_cache *bh_cachep __read_mostly;
2868
2869/*
2870 * Once the number of bh's in the machine exceeds this level, we start
2871 * stripping them in writeback.
2872 */
2873static unsigned long max_buffer_heads;
2874
2875int buffer_heads_over_limit;
2876
2877struct bh_accounting {
2878 int nr; /* Number of live bh's */
2879 int ratelimit; /* Limit cacheline bouncing */
2880};
2881
2882static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2883
2884static void recalc_bh_state(void)
2885{
2886 int i;
2887 int tot = 0;
2888
2889 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
2890 return;
2891 __this_cpu_write(bh_accounting.ratelimit, 0);
2892 for_each_online_cpu(i)
2893 tot += per_cpu(bh_accounting, i).nr;
2894 buffer_heads_over_limit = (tot > max_buffer_heads);
2895}
2896
2897struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2898{
2899 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
2900 if (ret) {
2901 INIT_LIST_HEAD(&ret->b_assoc_buffers);
2902 spin_lock_init(&ret->b_uptodate_lock);
2903 preempt_disable();
2904 __this_cpu_inc(bh_accounting.nr);
2905 recalc_bh_state();
2906 preempt_enable();
2907 }
2908 return ret;
2909}
2910EXPORT_SYMBOL(alloc_buffer_head);
2911
2912void free_buffer_head(struct buffer_head *bh)
2913{
2914 BUG_ON(!list_empty(&bh->b_assoc_buffers));
2915 kmem_cache_free(bh_cachep, bh);
2916 preempt_disable();
2917 __this_cpu_dec(bh_accounting.nr);
2918 recalc_bh_state();
2919 preempt_enable();
2920}
2921EXPORT_SYMBOL(free_buffer_head);
2922
2923static int buffer_exit_cpu_dead(unsigned int cpu)
2924{
2925 int i;
2926 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2927
2928 for (i = 0; i < BH_LRU_SIZE; i++) {
2929 brelse(b->bhs[i]);
2930 b->bhs[i] = NULL;
2931 }
2932 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
2933 per_cpu(bh_accounting, cpu).nr = 0;
2934 return 0;
2935}
2936
2937/**
2938 * bh_uptodate_or_lock - Test whether the buffer is uptodate
2939 * @bh: struct buffer_head
2940 *
2941 * Return true if the buffer is up-to-date and false,
2942 * with the buffer locked, if not.
2943 */
2944int bh_uptodate_or_lock(struct buffer_head *bh)
2945{
2946 if (!buffer_uptodate(bh)) {
2947 lock_buffer(bh);
2948 if (!buffer_uptodate(bh))
2949 return 0;
2950 unlock_buffer(bh);
2951 }
2952 return 1;
2953}
2954EXPORT_SYMBOL(bh_uptodate_or_lock);
2955
2956/**
2957 * __bh_read - Submit read for a locked buffer
2958 * @bh: struct buffer_head
2959 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
2960 * @wait: wait until reading finish
2961 *
2962 * Returns zero on success or don't wait, and -EIO on error.
2963 */
2964int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
2965{
2966 int ret = 0;
2967
2968 BUG_ON(!buffer_locked(bh));
2969
2970 get_bh(bh);
2971 bh->b_end_io = end_buffer_read_sync;
2972 submit_bh(REQ_OP_READ | op_flags, bh);
2973 if (wait) {
2974 wait_on_buffer(bh);
2975 if (!buffer_uptodate(bh))
2976 ret = -EIO;
2977 }
2978 return ret;
2979}
2980EXPORT_SYMBOL(__bh_read);
2981
2982/**
2983 * __bh_read_batch - Submit read for a batch of unlocked buffers
2984 * @nr: entry number of the buffer batch
2985 * @bhs: a batch of struct buffer_head
2986 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
2987 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
2988 * buffer that cannot lock.
2989 *
2990 * Returns zero on success or don't wait, and -EIO on error.
2991 */
2992void __bh_read_batch(int nr, struct buffer_head *bhs[],
2993 blk_opf_t op_flags, bool force_lock)
2994{
2995 int i;
2996
2997 for (i = 0; i < nr; i++) {
2998 struct buffer_head *bh = bhs[i];
2999
3000 if (buffer_uptodate(bh))
3001 continue;
3002
3003 if (force_lock)
3004 lock_buffer(bh);
3005 else
3006 if (!trylock_buffer(bh))
3007 continue;
3008
3009 if (buffer_uptodate(bh)) {
3010 unlock_buffer(bh);
3011 continue;
3012 }
3013
3014 bh->b_end_io = end_buffer_read_sync;
3015 get_bh(bh);
3016 submit_bh(REQ_OP_READ | op_flags, bh);
3017 }
3018}
3019EXPORT_SYMBOL(__bh_read_batch);
3020
3021void __init buffer_init(void)
3022{
3023 unsigned long nrpages;
3024 int ret;
3025
3026 bh_cachep = kmem_cache_create("buffer_head",
3027 sizeof(struct buffer_head), 0,
3028 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3029 SLAB_MEM_SPREAD),
3030 NULL);
3031
3032 /*
3033 * Limit the bh occupancy to 10% of ZONE_NORMAL
3034 */
3035 nrpages = (nr_free_buffer_pages() * 10) / 100;
3036 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3037 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3038 NULL, buffer_exit_cpu_dead);
3039 WARN_ON(ret < 0);
3040}