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
51static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
52static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
53 enum rw_hint hint, struct writeback_control *wbc);
54
55#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
56
57inline void touch_buffer(struct buffer_head *bh)
58{
59 trace_block_touch_buffer(bh);
60 mark_page_accessed(bh->b_page);
61}
62EXPORT_SYMBOL(touch_buffer);
63
64void __lock_buffer(struct buffer_head *bh)
65{
66 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
67}
68EXPORT_SYMBOL(__lock_buffer);
69
70void unlock_buffer(struct buffer_head *bh)
71{
72 clear_bit_unlock(BH_Lock, &bh->b_state);
73 smp_mb__after_atomic();
74 wake_up_bit(&bh->b_state, BH_Lock);
75}
76EXPORT_SYMBOL(unlock_buffer);
77
78/*
79 * Returns if the page has dirty or writeback buffers. If all the buffers
80 * are unlocked and clean then the PageDirty information is stale. If
81 * any of the pages are locked, it is assumed they are locked for IO.
82 */
83void buffer_check_dirty_writeback(struct page *page,
84 bool *dirty, bool *writeback)
85{
86 struct buffer_head *head, *bh;
87 *dirty = false;
88 *writeback = false;
89
90 BUG_ON(!PageLocked(page));
91
92 if (!page_has_buffers(page))
93 return;
94
95 if (PageWriteback(page))
96 *writeback = true;
97
98 head = page_buffers(page);
99 bh = head;
100 do {
101 if (buffer_locked(bh))
102 *writeback = true;
103
104 if (buffer_dirty(bh))
105 *dirty = true;
106
107 bh = bh->b_this_page;
108 } while (bh != head);
109}
110EXPORT_SYMBOL(buffer_check_dirty_writeback);
111
112/*
113 * Block until a buffer comes unlocked. This doesn't stop it
114 * from becoming locked again - you have to lock it yourself
115 * if you want to preserve its state.
116 */
117void __wait_on_buffer(struct buffer_head * bh)
118{
119 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
120}
121EXPORT_SYMBOL(__wait_on_buffer);
122
123static void
124__clear_page_buffers(struct page *page)
125{
126 ClearPagePrivate(page);
127 set_page_private(page, 0);
128 put_page(page);
129}
130
131static void buffer_io_error(struct buffer_head *bh, char *msg)
132{
133 if (!test_bit(BH_Quiet, &bh->b_state))
134 printk_ratelimited(KERN_ERR
135 "Buffer I/O error on dev %pg, logical block %llu%s\n",
136 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
137}
138
139/*
140 * End-of-IO handler helper function which does not touch the bh after
141 * unlocking it.
142 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
143 * a race there is benign: unlock_buffer() only use the bh's address for
144 * hashing after unlocking the buffer, so it doesn't actually touch the bh
145 * itself.
146 */
147static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
148{
149 if (uptodate) {
150 set_buffer_uptodate(bh);
151 } else {
152 /* This happens, due to failed read-ahead attempts. */
153 clear_buffer_uptodate(bh);
154 }
155 unlock_buffer(bh);
156}
157
158/*
159 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
160 * unlock the buffer. This is what ll_rw_block uses too.
161 */
162void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
163{
164 __end_buffer_read_notouch(bh, uptodate);
165 put_bh(bh);
166}
167EXPORT_SYMBOL(end_buffer_read_sync);
168
169void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
170{
171 if (uptodate) {
172 set_buffer_uptodate(bh);
173 } else {
174 buffer_io_error(bh, ", lost sync page write");
175 mark_buffer_write_io_error(bh);
176 clear_buffer_uptodate(bh);
177 }
178 unlock_buffer(bh);
179 put_bh(bh);
180}
181EXPORT_SYMBOL(end_buffer_write_sync);
182
183/*
184 * Various filesystems appear to want __find_get_block to be non-blocking.
185 * But it's the page lock which protects the buffers. To get around this,
186 * we get exclusion from try_to_free_buffers with the blockdev mapping's
187 * private_lock.
188 *
189 * Hack idea: for the blockdev mapping, private_lock contention
190 * may be quite high. This code could TryLock the page, and if that
191 * succeeds, there is no need to take private_lock.
192 */
193static struct buffer_head *
194__find_get_block_slow(struct block_device *bdev, sector_t block)
195{
196 struct inode *bd_inode = bdev->bd_inode;
197 struct address_space *bd_mapping = bd_inode->i_mapping;
198 struct buffer_head *ret = NULL;
199 pgoff_t index;
200 struct buffer_head *bh;
201 struct buffer_head *head;
202 struct page *page;
203 int all_mapped = 1;
204 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
205
206 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
207 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
208 if (!page)
209 goto out;
210
211 spin_lock(&bd_mapping->private_lock);
212 if (!page_has_buffers(page))
213 goto out_unlock;
214 head = page_buffers(page);
215 bh = head;
216 do {
217 if (!buffer_mapped(bh))
218 all_mapped = 0;
219 else if (bh->b_blocknr == block) {
220 ret = bh;
221 get_bh(bh);
222 goto out_unlock;
223 }
224 bh = bh->b_this_page;
225 } while (bh != head);
226
227 /* we might be here because some of the buffers on this page are
228 * not mapped. This is due to various races between
229 * file io on the block device and getblk. It gets dealt with
230 * elsewhere, don't buffer_error if we had some unmapped buffers
231 */
232 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
233 if (all_mapped && __ratelimit(&last_warned)) {
234 printk("__find_get_block_slow() failed. block=%llu, "
235 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
236 "device %pg blocksize: %d\n",
237 (unsigned long long)block,
238 (unsigned long long)bh->b_blocknr,
239 bh->b_state, bh->b_size, bdev,
240 1 << bd_inode->i_blkbits);
241 }
242out_unlock:
243 spin_unlock(&bd_mapping->private_lock);
244 put_page(page);
245out:
246 return ret;
247}
248
249/*
250 * I/O completion handler for block_read_full_page() - pages
251 * which come unlocked at the end of I/O.
252 */
253static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
254{
255 unsigned long flags;
256 struct buffer_head *first;
257 struct buffer_head *tmp;
258 struct page *page;
259 int page_uptodate = 1;
260
261 BUG_ON(!buffer_async_read(bh));
262
263 page = bh->b_page;
264 if (uptodate) {
265 set_buffer_uptodate(bh);
266 } else {
267 clear_buffer_uptodate(bh);
268 buffer_io_error(bh, ", async page read");
269 SetPageError(page);
270 }
271
272 /*
273 * Be _very_ careful from here on. Bad things can happen if
274 * two buffer heads end IO at almost the same time and both
275 * decide that the page is now completely done.
276 */
277 first = page_buffers(page);
278 local_irq_save(flags);
279 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
280 clear_buffer_async_read(bh);
281 unlock_buffer(bh);
282 tmp = bh;
283 do {
284 if (!buffer_uptodate(tmp))
285 page_uptodate = 0;
286 if (buffer_async_read(tmp)) {
287 BUG_ON(!buffer_locked(tmp));
288 goto still_busy;
289 }
290 tmp = tmp->b_this_page;
291 } while (tmp != bh);
292 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
293 local_irq_restore(flags);
294
295 /*
296 * If none of the buffers had errors and they are all
297 * uptodate then we can set the page uptodate.
298 */
299 if (page_uptodate && !PageError(page))
300 SetPageUptodate(page);
301 unlock_page(page);
302 return;
303
304still_busy:
305 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
306 local_irq_restore(flags);
307 return;
308}
309
310/*
311 * Completion handler for block_write_full_page() - pages which are unlocked
312 * during I/O, and which have PageWriteback cleared upon I/O completion.
313 */
314void end_buffer_async_write(struct buffer_head *bh, int uptodate)
315{
316 unsigned long flags;
317 struct buffer_head *first;
318 struct buffer_head *tmp;
319 struct page *page;
320
321 BUG_ON(!buffer_async_write(bh));
322
323 page = bh->b_page;
324 if (uptodate) {
325 set_buffer_uptodate(bh);
326 } else {
327 buffer_io_error(bh, ", lost async page write");
328 mark_buffer_write_io_error(bh);
329 clear_buffer_uptodate(bh);
330 SetPageError(page);
331 }
332
333 first = page_buffers(page);
334 local_irq_save(flags);
335 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336
337 clear_buffer_async_write(bh);
338 unlock_buffer(bh);
339 tmp = bh->b_this_page;
340 while (tmp != bh) {
341 if (buffer_async_write(tmp)) {
342 BUG_ON(!buffer_locked(tmp));
343 goto still_busy;
344 }
345 tmp = tmp->b_this_page;
346 }
347 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
348 local_irq_restore(flags);
349 end_page_writeback(page);
350 return;
351
352still_busy:
353 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
354 local_irq_restore(flags);
355 return;
356}
357EXPORT_SYMBOL(end_buffer_async_write);
358
359/*
360 * If a page's buffers are under async readin (end_buffer_async_read
361 * completion) then there is a possibility that another thread of
362 * control could lock one of the buffers after it has completed
363 * but while some of the other buffers have not completed. This
364 * locked buffer would confuse end_buffer_async_read() into not unlocking
365 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
366 * that this buffer is not under async I/O.
367 *
368 * The page comes unlocked when it has no locked buffer_async buffers
369 * left.
370 *
371 * PageLocked prevents anyone starting new async I/O reads any of
372 * the buffers.
373 *
374 * PageWriteback is used to prevent simultaneous writeout of the same
375 * page.
376 *
377 * PageLocked prevents anyone from starting writeback of a page which is
378 * under read I/O (PageWriteback is only ever set against a locked page).
379 */
380static void mark_buffer_async_read(struct buffer_head *bh)
381{
382 bh->b_end_io = end_buffer_async_read;
383 set_buffer_async_read(bh);
384}
385
386static void mark_buffer_async_write_endio(struct buffer_head *bh,
387 bh_end_io_t *handler)
388{
389 bh->b_end_io = handler;
390 set_buffer_async_write(bh);
391}
392
393void mark_buffer_async_write(struct buffer_head *bh)
394{
395 mark_buffer_async_write_endio(bh, end_buffer_async_write);
396}
397EXPORT_SYMBOL(mark_buffer_async_write);
398
399
400/*
401 * fs/buffer.c contains helper functions for buffer-backed address space's
402 * fsync functions. A common requirement for buffer-based filesystems is
403 * that certain data from the backing blockdev needs to be written out for
404 * a successful fsync(). For example, ext2 indirect blocks need to be
405 * written back and waited upon before fsync() returns.
406 *
407 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
408 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
409 * management of a list of dependent buffers at ->i_mapping->private_list.
410 *
411 * Locking is a little subtle: try_to_free_buffers() will remove buffers
412 * from their controlling inode's queue when they are being freed. But
413 * try_to_free_buffers() will be operating against the *blockdev* mapping
414 * at the time, not against the S_ISREG file which depends on those buffers.
415 * So the locking for private_list is via the private_lock in the address_space
416 * which backs the buffers. Which is different from the address_space
417 * against which the buffers are listed. So for a particular address_space,
418 * mapping->private_lock does *not* protect mapping->private_list! In fact,
419 * mapping->private_list will always be protected by the backing blockdev's
420 * ->private_lock.
421 *
422 * Which introduces a requirement: all buffers on an address_space's
423 * ->private_list must be from the same address_space: the blockdev's.
424 *
425 * address_spaces which do not place buffers at ->private_list via these
426 * utility functions are free to use private_lock and private_list for
427 * whatever they want. The only requirement is that list_empty(private_list)
428 * be true at clear_inode() time.
429 *
430 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
431 * filesystems should do that. invalidate_inode_buffers() should just go
432 * BUG_ON(!list_empty).
433 *
434 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
435 * take an address_space, not an inode. And it should be called
436 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
437 * queued up.
438 *
439 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
440 * list if it is already on a list. Because if the buffer is on a list,
441 * it *must* already be on the right one. If not, the filesystem is being
442 * silly. This will save a ton of locking. But first we have to ensure
443 * that buffers are taken *off* the old inode's list when they are freed
444 * (presumably in truncate). That requires careful auditing of all
445 * filesystems (do it inside bforget()). It could also be done by bringing
446 * b_inode back.
447 */
448
449/*
450 * The buffer's backing address_space's private_lock must be held
451 */
452static void __remove_assoc_queue(struct buffer_head *bh)
453{
454 list_del_init(&bh->b_assoc_buffers);
455 WARN_ON(!bh->b_assoc_map);
456 bh->b_assoc_map = NULL;
457}
458
459int inode_has_buffers(struct inode *inode)
460{
461 return !list_empty(&inode->i_data.private_list);
462}
463
464/*
465 * osync is designed to support O_SYNC io. It waits synchronously for
466 * all already-submitted IO to complete, but does not queue any new
467 * writes to the disk.
468 *
469 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
470 * you dirty the buffers, and then use osync_inode_buffers to wait for
471 * completion. Any other dirty buffers which are not yet queued for
472 * write will not be flushed to disk by the osync.
473 */
474static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
475{
476 struct buffer_head *bh;
477 struct list_head *p;
478 int err = 0;
479
480 spin_lock(lock);
481repeat:
482 list_for_each_prev(p, list) {
483 bh = BH_ENTRY(p);
484 if (buffer_locked(bh)) {
485 get_bh(bh);
486 spin_unlock(lock);
487 wait_on_buffer(bh);
488 if (!buffer_uptodate(bh))
489 err = -EIO;
490 brelse(bh);
491 spin_lock(lock);
492 goto repeat;
493 }
494 }
495 spin_unlock(lock);
496 return err;
497}
498
499void emergency_thaw_bdev(struct super_block *sb)
500{
501 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
502 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
503}
504
505/**
506 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
507 * @mapping: the mapping which wants those buffers written
508 *
509 * Starts I/O against the buffers at mapping->private_list, and waits upon
510 * that I/O.
511 *
512 * Basically, this is a convenience function for fsync().
513 * @mapping is a file or directory which needs those buffers to be written for
514 * a successful fsync().
515 */
516int sync_mapping_buffers(struct address_space *mapping)
517{
518 struct address_space *buffer_mapping = mapping->private_data;
519
520 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
521 return 0;
522
523 return fsync_buffers_list(&buffer_mapping->private_lock,
524 &mapping->private_list);
525}
526EXPORT_SYMBOL(sync_mapping_buffers);
527
528/*
529 * Called when we've recently written block `bblock', and it is known that
530 * `bblock' was for a buffer_boundary() buffer. This means that the block at
531 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
532 * dirty, schedule it for IO. So that indirects merge nicely with their data.
533 */
534void write_boundary_block(struct block_device *bdev,
535 sector_t bblock, unsigned blocksize)
536{
537 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
538 if (bh) {
539 if (buffer_dirty(bh))
540 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
541 put_bh(bh);
542 }
543}
544
545void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
546{
547 struct address_space *mapping = inode->i_mapping;
548 struct address_space *buffer_mapping = bh->b_page->mapping;
549
550 mark_buffer_dirty(bh);
551 if (!mapping->private_data) {
552 mapping->private_data = buffer_mapping;
553 } else {
554 BUG_ON(mapping->private_data != buffer_mapping);
555 }
556 if (!bh->b_assoc_map) {
557 spin_lock(&buffer_mapping->private_lock);
558 list_move_tail(&bh->b_assoc_buffers,
559 &mapping->private_list);
560 bh->b_assoc_map = mapping;
561 spin_unlock(&buffer_mapping->private_lock);
562 }
563}
564EXPORT_SYMBOL(mark_buffer_dirty_inode);
565
566/*
567 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
568 * dirty.
569 *
570 * If warn is true, then emit a warning if the page is not uptodate and has
571 * not been truncated.
572 *
573 * The caller must hold lock_page_memcg().
574 */
575void __set_page_dirty(struct page *page, struct address_space *mapping,
576 int warn)
577{
578 unsigned long flags;
579
580 xa_lock_irqsave(&mapping->i_pages, flags);
581 if (page->mapping) { /* Race with truncate? */
582 WARN_ON_ONCE(warn && !PageUptodate(page));
583 account_page_dirtied(page, mapping);
584 __xa_set_mark(&mapping->i_pages, page_index(page),
585 PAGECACHE_TAG_DIRTY);
586 }
587 xa_unlock_irqrestore(&mapping->i_pages, flags);
588}
589EXPORT_SYMBOL_GPL(__set_page_dirty);
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 */
616int __set_page_dirty_buffers(struct page *page)
617{
618 int newly_dirty;
619 struct address_space *mapping = page_mapping(page);
620
621 if (unlikely(!mapping))
622 return !TestSetPageDirty(page);
623
624 spin_lock(&mapping->private_lock);
625 if (page_has_buffers(page)) {
626 struct buffer_head *head = page_buffers(page);
627 struct buffer_head *bh = head;
628
629 do {
630 set_buffer_dirty(bh);
631 bh = bh->b_this_page;
632 } while (bh != head);
633 }
634 /*
635 * Lock out page->mem_cgroup migration to keep PageDirty
636 * synchronized with per-memcg dirty page counters.
637 */
638 lock_page_memcg(page);
639 newly_dirty = !TestSetPageDirty(page);
640 spin_unlock(&mapping->private_lock);
641
642 if (newly_dirty)
643 __set_page_dirty(page, mapping, 1);
644
645 unlock_page_memcg(page);
646
647 if (newly_dirty)
648 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
649
650 return newly_dirty;
651}
652EXPORT_SYMBOL(__set_page_dirty_buffers);
653
654/*
655 * Write out and wait upon a list of buffers.
656 *
657 * We have conflicting pressures: we want to make sure that all
658 * initially dirty buffers get waited on, but that any subsequently
659 * dirtied buffers don't. After all, we don't want fsync to last
660 * forever if somebody is actively writing to the file.
661 *
662 * Do this in two main stages: first we copy dirty buffers to a
663 * temporary inode list, queueing the writes as we go. Then we clean
664 * up, waiting for those writes to complete.
665 *
666 * During this second stage, any subsequent updates to the file may end
667 * up refiling the buffer on the original inode's dirty list again, so
668 * there is a chance we will end up with a buffer queued for write but
669 * not yet completed on that list. So, as a final cleanup we go through
670 * the osync code to catch these locked, dirty buffers without requeuing
671 * any newly dirty buffers for write.
672 */
673static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
674{
675 struct buffer_head *bh;
676 struct list_head tmp;
677 struct address_space *mapping;
678 int err = 0, err2;
679 struct blk_plug plug;
680
681 INIT_LIST_HEAD(&tmp);
682 blk_start_plug(&plug);
683
684 spin_lock(lock);
685 while (!list_empty(list)) {
686 bh = BH_ENTRY(list->next);
687 mapping = bh->b_assoc_map;
688 __remove_assoc_queue(bh);
689 /* Avoid race with mark_buffer_dirty_inode() which does
690 * a lockless check and we rely on seeing the dirty bit */
691 smp_mb();
692 if (buffer_dirty(bh) || buffer_locked(bh)) {
693 list_add(&bh->b_assoc_buffers, &tmp);
694 bh->b_assoc_map = mapping;
695 if (buffer_dirty(bh)) {
696 get_bh(bh);
697 spin_unlock(lock);
698 /*
699 * Ensure any pending I/O completes so that
700 * write_dirty_buffer() actually writes the
701 * current contents - it is a noop if I/O is
702 * still in flight on potentially older
703 * contents.
704 */
705 write_dirty_buffer(bh, REQ_SYNC);
706
707 /*
708 * Kick off IO for the previous mapping. Note
709 * that we will not run the very last mapping,
710 * wait_on_buffer() will do that for us
711 * through sync_buffer().
712 */
713 brelse(bh);
714 spin_lock(lock);
715 }
716 }
717 }
718
719 spin_unlock(lock);
720 blk_finish_plug(&plug);
721 spin_lock(lock);
722
723 while (!list_empty(&tmp)) {
724 bh = BH_ENTRY(tmp.prev);
725 get_bh(bh);
726 mapping = bh->b_assoc_map;
727 __remove_assoc_queue(bh);
728 /* Avoid race with mark_buffer_dirty_inode() which does
729 * a lockless check and we rely on seeing the dirty bit */
730 smp_mb();
731 if (buffer_dirty(bh)) {
732 list_add(&bh->b_assoc_buffers,
733 &mapping->private_list);
734 bh->b_assoc_map = mapping;
735 }
736 spin_unlock(lock);
737 wait_on_buffer(bh);
738 if (!buffer_uptodate(bh))
739 err = -EIO;
740 brelse(bh);
741 spin_lock(lock);
742 }
743
744 spin_unlock(lock);
745 err2 = osync_buffers_list(lock, list);
746 if (err)
747 return err;
748 else
749 return err2;
750}
751
752/*
753 * Invalidate any and all dirty buffers on a given inode. We are
754 * probably unmounting the fs, but that doesn't mean we have already
755 * done a sync(). Just drop the buffers from the inode list.
756 *
757 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
758 * assumes that all the buffers are against the blockdev. Not true
759 * for reiserfs.
760 */
761void invalidate_inode_buffers(struct inode *inode)
762{
763 if (inode_has_buffers(inode)) {
764 struct address_space *mapping = &inode->i_data;
765 struct list_head *list = &mapping->private_list;
766 struct address_space *buffer_mapping = mapping->private_data;
767
768 spin_lock(&buffer_mapping->private_lock);
769 while (!list_empty(list))
770 __remove_assoc_queue(BH_ENTRY(list->next));
771 spin_unlock(&buffer_mapping->private_lock);
772 }
773}
774EXPORT_SYMBOL(invalidate_inode_buffers);
775
776/*
777 * Remove any clean buffers from the inode's buffer list. This is called
778 * when we're trying to free the inode itself. Those buffers can pin it.
779 *
780 * Returns true if all buffers were removed.
781 */
782int remove_inode_buffers(struct inode *inode)
783{
784 int ret = 1;
785
786 if (inode_has_buffers(inode)) {
787 struct address_space *mapping = &inode->i_data;
788 struct list_head *list = &mapping->private_list;
789 struct address_space *buffer_mapping = mapping->private_data;
790
791 spin_lock(&buffer_mapping->private_lock);
792 while (!list_empty(list)) {
793 struct buffer_head *bh = BH_ENTRY(list->next);
794 if (buffer_dirty(bh)) {
795 ret = 0;
796 break;
797 }
798 __remove_assoc_queue(bh);
799 }
800 spin_unlock(&buffer_mapping->private_lock);
801 }
802 return ret;
803}
804
805/*
806 * Create the appropriate buffers when given a page for data area and
807 * the size of each buffer.. Use the bh->b_this_page linked list to
808 * follow the buffers created. Return NULL if unable to create more
809 * buffers.
810 *
811 * The retry flag is used to differentiate async IO (paging, swapping)
812 * which may not fail from ordinary buffer allocations.
813 */
814struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
815 bool retry)
816{
817 struct buffer_head *bh, *head;
818 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
819 long offset;
820 struct mem_cgroup *memcg;
821
822 if (retry)
823 gfp |= __GFP_NOFAIL;
824
825 memcg = get_mem_cgroup_from_page(page);
826 memalloc_use_memcg(memcg);
827
828 head = NULL;
829 offset = PAGE_SIZE;
830 while ((offset -= size) >= 0) {
831 bh = alloc_buffer_head(gfp);
832 if (!bh)
833 goto no_grow;
834
835 bh->b_this_page = head;
836 bh->b_blocknr = -1;
837 head = bh;
838
839 bh->b_size = size;
840
841 /* Link the buffer to its page */
842 set_bh_page(bh, page, offset);
843 }
844out:
845 memalloc_unuse_memcg();
846 mem_cgroup_put(memcg);
847 return head;
848/*
849 * In case anything failed, we just free everything we got.
850 */
851no_grow:
852 if (head) {
853 do {
854 bh = head;
855 head = head->b_this_page;
856 free_buffer_head(bh);
857 } while (head);
858 }
859
860 goto out;
861}
862EXPORT_SYMBOL_GPL(alloc_page_buffers);
863
864static inline void
865link_dev_buffers(struct page *page, struct buffer_head *head)
866{
867 struct buffer_head *bh, *tail;
868
869 bh = head;
870 do {
871 tail = bh;
872 bh = bh->b_this_page;
873 } while (bh);
874 tail->b_this_page = head;
875 attach_page_buffers(page, head);
876}
877
878static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
879{
880 sector_t retval = ~((sector_t)0);
881 loff_t sz = i_size_read(bdev->bd_inode);
882
883 if (sz) {
884 unsigned int sizebits = blksize_bits(size);
885 retval = (sz >> sizebits);
886 }
887 return retval;
888}
889
890/*
891 * Initialise the state of a blockdev page's buffers.
892 */
893static sector_t
894init_page_buffers(struct page *page, struct block_device *bdev,
895 sector_t block, int size)
896{
897 struct buffer_head *head = page_buffers(page);
898 struct buffer_head *bh = head;
899 int uptodate = PageUptodate(page);
900 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
901
902 do {
903 if (!buffer_mapped(bh)) {
904 bh->b_end_io = NULL;
905 bh->b_private = NULL;
906 bh->b_bdev = bdev;
907 bh->b_blocknr = block;
908 if (uptodate)
909 set_buffer_uptodate(bh);
910 if (block < end_block)
911 set_buffer_mapped(bh);
912 }
913 block++;
914 bh = bh->b_this_page;
915 } while (bh != head);
916
917 /*
918 * Caller needs to validate requested block against end of device.
919 */
920 return end_block;
921}
922
923/*
924 * Create the page-cache page that contains the requested block.
925 *
926 * This is used purely for blockdev mappings.
927 */
928static int
929grow_dev_page(struct block_device *bdev, sector_t block,
930 pgoff_t index, int size, int sizebits, gfp_t gfp)
931{
932 struct inode *inode = bdev->bd_inode;
933 struct page *page;
934 struct buffer_head *bh;
935 sector_t end_block;
936 int ret = 0; /* Will call free_more_memory() */
937 gfp_t gfp_mask;
938
939 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
940
941 /*
942 * XXX: __getblk_slow() can not really deal with failure and
943 * will endlessly loop on improvised global reclaim. Prefer
944 * looping in the allocator rather than here, at least that
945 * code knows what it's doing.
946 */
947 gfp_mask |= __GFP_NOFAIL;
948
949 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
950
951 BUG_ON(!PageLocked(page));
952
953 if (page_has_buffers(page)) {
954 bh = page_buffers(page);
955 if (bh->b_size == size) {
956 end_block = init_page_buffers(page, bdev,
957 (sector_t)index << sizebits,
958 size);
959 goto done;
960 }
961 if (!try_to_free_buffers(page))
962 goto failed;
963 }
964
965 /*
966 * Allocate some buffers for this page
967 */
968 bh = alloc_page_buffers(page, size, true);
969
970 /*
971 * Link the page to the buffers and initialise them. Take the
972 * lock to be atomic wrt __find_get_block(), which does not
973 * run under the page lock.
974 */
975 spin_lock(&inode->i_mapping->private_lock);
976 link_dev_buffers(page, bh);
977 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
978 size);
979 spin_unlock(&inode->i_mapping->private_lock);
980done:
981 ret = (block < end_block) ? 1 : -ENXIO;
982failed:
983 unlock_page(page);
984 put_page(page);
985 return ret;
986}
987
988/*
989 * Create buffers for the specified block device block's page. If
990 * that page was dirty, the buffers are set dirty also.
991 */
992static int
993grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
994{
995 pgoff_t index;
996 int sizebits;
997
998 sizebits = -1;
999 do {
1000 sizebits++;
1001 } while ((size << sizebits) < PAGE_SIZE);
1002
1003 index = block >> sizebits;
1004
1005 /*
1006 * Check for a block which wants to lie outside our maximum possible
1007 * pagecache index. (this comparison is done using sector_t types).
1008 */
1009 if (unlikely(index != block >> sizebits)) {
1010 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1011 "device %pg\n",
1012 __func__, (unsigned long long)block,
1013 bdev);
1014 return -EIO;
1015 }
1016
1017 /* Create a page with the proper size buffers.. */
1018 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1019}
1020
1021static struct buffer_head *
1022__getblk_slow(struct block_device *bdev, sector_t block,
1023 unsigned size, gfp_t gfp)
1024{
1025 /* Size must be multiple of hard sectorsize */
1026 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1027 (size < 512 || size > PAGE_SIZE))) {
1028 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1029 size);
1030 printk(KERN_ERR "logical block size: %d\n",
1031 bdev_logical_block_size(bdev));
1032
1033 dump_stack();
1034 return NULL;
1035 }
1036
1037 for (;;) {
1038 struct buffer_head *bh;
1039 int ret;
1040
1041 bh = __find_get_block(bdev, block, size);
1042 if (bh)
1043 return bh;
1044
1045 ret = grow_buffers(bdev, block, size, gfp);
1046 if (ret < 0)
1047 return NULL;
1048 }
1049}
1050
1051/*
1052 * The relationship between dirty buffers and dirty pages:
1053 *
1054 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1055 * the page is tagged dirty in the page cache.
1056 *
1057 * At all times, the dirtiness of the buffers represents the dirtiness of
1058 * subsections of the page. If the page has buffers, the page dirty bit is
1059 * merely a hint about the true dirty state.
1060 *
1061 * When a page is set dirty in its entirety, all its buffers are marked dirty
1062 * (if the page has buffers).
1063 *
1064 * When a buffer is marked dirty, its page is dirtied, but the page's other
1065 * buffers are not.
1066 *
1067 * Also. When blockdev buffers are explicitly read with bread(), they
1068 * individually become uptodate. But their backing page remains not
1069 * uptodate - even if all of its buffers are uptodate. A subsequent
1070 * block_read_full_page() against that page will discover all the uptodate
1071 * buffers, will set the page uptodate and will perform no I/O.
1072 */
1073
1074/**
1075 * mark_buffer_dirty - mark a buffer_head as needing writeout
1076 * @bh: the buffer_head to mark dirty
1077 *
1078 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1079 * its backing page dirty, then tag the page as dirty in the page cache
1080 * and then attach the address_space's inode to its superblock's dirty
1081 * inode list.
1082 *
1083 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1084 * i_pages lock and mapping->host->i_lock.
1085 */
1086void mark_buffer_dirty(struct buffer_head *bh)
1087{
1088 WARN_ON_ONCE(!buffer_uptodate(bh));
1089
1090 trace_block_dirty_buffer(bh);
1091
1092 /*
1093 * Very *carefully* optimize the it-is-already-dirty case.
1094 *
1095 * Don't let the final "is it dirty" escape to before we
1096 * perhaps modified the buffer.
1097 */
1098 if (buffer_dirty(bh)) {
1099 smp_mb();
1100 if (buffer_dirty(bh))
1101 return;
1102 }
1103
1104 if (!test_set_buffer_dirty(bh)) {
1105 struct page *page = bh->b_page;
1106 struct address_space *mapping = NULL;
1107
1108 lock_page_memcg(page);
1109 if (!TestSetPageDirty(page)) {
1110 mapping = page_mapping(page);
1111 if (mapping)
1112 __set_page_dirty(page, mapping, 0);
1113 }
1114 unlock_page_memcg(page);
1115 if (mapping)
1116 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1117 }
1118}
1119EXPORT_SYMBOL(mark_buffer_dirty);
1120
1121void mark_buffer_write_io_error(struct buffer_head *bh)
1122{
1123 set_buffer_write_io_error(bh);
1124 /* FIXME: do we need to set this in both places? */
1125 if (bh->b_page && bh->b_page->mapping)
1126 mapping_set_error(bh->b_page->mapping, -EIO);
1127 if (bh->b_assoc_map)
1128 mapping_set_error(bh->b_assoc_map, -EIO);
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, 0, 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 b = this_cpu_ptr(&bh_lrus);
1238 for (i = 0; i < BH_LRU_SIZE; i++) {
1239 swap(evictee, b->bhs[i]);
1240 if (evictee == bh) {
1241 bh_lru_unlock();
1242 return;
1243 }
1244 }
1245
1246 get_bh(bh);
1247 bh_lru_unlock();
1248 brelse(evictee);
1249}
1250
1251/*
1252 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1253 */
1254static struct buffer_head *
1255lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1256{
1257 struct buffer_head *ret = NULL;
1258 unsigned int i;
1259
1260 check_irqs_on();
1261 bh_lru_lock();
1262 for (i = 0; i < BH_LRU_SIZE; i++) {
1263 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1264
1265 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1266 bh->b_size == size) {
1267 if (i) {
1268 while (i) {
1269 __this_cpu_write(bh_lrus.bhs[i],
1270 __this_cpu_read(bh_lrus.bhs[i - 1]));
1271 i--;
1272 }
1273 __this_cpu_write(bh_lrus.bhs[0], bh);
1274 }
1275 get_bh(bh);
1276 ret = bh;
1277 break;
1278 }
1279 }
1280 bh_lru_unlock();
1281 return ret;
1282}
1283
1284/*
1285 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1286 * it in the LRU and mark it as accessed. If it is not present then return
1287 * NULL
1288 */
1289struct buffer_head *
1290__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1291{
1292 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1293
1294 if (bh == NULL) {
1295 /* __find_get_block_slow will mark the page accessed */
1296 bh = __find_get_block_slow(bdev, block);
1297 if (bh)
1298 bh_lru_install(bh);
1299 } else
1300 touch_buffer(bh);
1301
1302 return bh;
1303}
1304EXPORT_SYMBOL(__find_get_block);
1305
1306/*
1307 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1308 * which corresponds to the passed block_device, block and size. The
1309 * returned buffer has its reference count incremented.
1310 *
1311 * __getblk_gfp() will lock up the machine if grow_dev_page's
1312 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1313 */
1314struct buffer_head *
1315__getblk_gfp(struct block_device *bdev, sector_t block,
1316 unsigned size, gfp_t gfp)
1317{
1318 struct buffer_head *bh = __find_get_block(bdev, block, size);
1319
1320 might_sleep();
1321 if (bh == NULL)
1322 bh = __getblk_slow(bdev, block, size, gfp);
1323 return bh;
1324}
1325EXPORT_SYMBOL(__getblk_gfp);
1326
1327/*
1328 * Do async read-ahead on a buffer..
1329 */
1330void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1331{
1332 struct buffer_head *bh = __getblk(bdev, block, size);
1333 if (likely(bh)) {
1334 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1335 brelse(bh);
1336 }
1337}
1338EXPORT_SYMBOL(__breadahead);
1339
1340/**
1341 * __bread_gfp() - reads a specified block and returns the bh
1342 * @bdev: the block_device to read from
1343 * @block: number of block
1344 * @size: size (in bytes) to read
1345 * @gfp: page allocation flag
1346 *
1347 * Reads a specified block, and returns buffer head that contains it.
1348 * The page cache can be allocated from non-movable area
1349 * not to prevent page migration if you set gfp to zero.
1350 * It returns NULL if the block was unreadable.
1351 */
1352struct buffer_head *
1353__bread_gfp(struct block_device *bdev, sector_t block,
1354 unsigned size, gfp_t gfp)
1355{
1356 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1357
1358 if (likely(bh) && !buffer_uptodate(bh))
1359 bh = __bread_slow(bh);
1360 return bh;
1361}
1362EXPORT_SYMBOL(__bread_gfp);
1363
1364/*
1365 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1366 * This doesn't race because it runs in each cpu either in irq
1367 * or with preempt disabled.
1368 */
1369static void invalidate_bh_lru(void *arg)
1370{
1371 struct bh_lru *b = &get_cpu_var(bh_lrus);
1372 int i;
1373
1374 for (i = 0; i < BH_LRU_SIZE; i++) {
1375 brelse(b->bhs[i]);
1376 b->bhs[i] = NULL;
1377 }
1378 put_cpu_var(bh_lrus);
1379}
1380
1381static bool has_bh_in_lru(int cpu, void *dummy)
1382{
1383 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1384 int i;
1385
1386 for (i = 0; i < BH_LRU_SIZE; i++) {
1387 if (b->bhs[i])
1388 return 1;
1389 }
1390
1391 return 0;
1392}
1393
1394void invalidate_bh_lrus(void)
1395{
1396 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1397}
1398EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1399
1400void set_bh_page(struct buffer_head *bh,
1401 struct page *page, unsigned long offset)
1402{
1403 bh->b_page = page;
1404 BUG_ON(offset >= PAGE_SIZE);
1405 if (PageHighMem(page))
1406 /*
1407 * This catches illegal uses and preserves the offset:
1408 */
1409 bh->b_data = (char *)(0 + offset);
1410 else
1411 bh->b_data = page_address(page) + offset;
1412}
1413EXPORT_SYMBOL(set_bh_page);
1414
1415/*
1416 * Called when truncating a buffer on a page completely.
1417 */
1418
1419/* Bits that are cleared during an invalidate */
1420#define BUFFER_FLAGS_DISCARD \
1421 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1422 1 << BH_Delay | 1 << BH_Unwritten)
1423
1424static void discard_buffer(struct buffer_head * bh)
1425{
1426 unsigned long b_state, b_state_old;
1427
1428 lock_buffer(bh);
1429 clear_buffer_dirty(bh);
1430 bh->b_bdev = NULL;
1431 b_state = bh->b_state;
1432 for (;;) {
1433 b_state_old = cmpxchg(&bh->b_state, b_state,
1434 (b_state & ~BUFFER_FLAGS_DISCARD));
1435 if (b_state_old == b_state)
1436 break;
1437 b_state = b_state_old;
1438 }
1439 unlock_buffer(bh);
1440}
1441
1442/**
1443 * block_invalidatepage - invalidate part or all of a buffer-backed page
1444 *
1445 * @page: the page which is affected
1446 * @offset: start of the range to invalidate
1447 * @length: length of the range to invalidate
1448 *
1449 * block_invalidatepage() is called when all or part of the page has become
1450 * invalidated by a truncate operation.
1451 *
1452 * block_invalidatepage() does not have to release all buffers, but it must
1453 * ensure that no dirty buffer is left outside @offset and that no I/O
1454 * is underway against any of the blocks which are outside the truncation
1455 * point. Because the caller is about to free (and possibly reuse) those
1456 * blocks on-disk.
1457 */
1458void block_invalidatepage(struct page *page, unsigned int offset,
1459 unsigned int length)
1460{
1461 struct buffer_head *head, *bh, *next;
1462 unsigned int curr_off = 0;
1463 unsigned int stop = length + offset;
1464
1465 BUG_ON(!PageLocked(page));
1466 if (!page_has_buffers(page))
1467 goto out;
1468
1469 /*
1470 * Check for overflow
1471 */
1472 BUG_ON(stop > PAGE_SIZE || stop < length);
1473
1474 head = page_buffers(page);
1475 bh = head;
1476 do {
1477 unsigned int next_off = curr_off + bh->b_size;
1478 next = bh->b_this_page;
1479
1480 /*
1481 * Are we still fully in range ?
1482 */
1483 if (next_off > stop)
1484 goto out;
1485
1486 /*
1487 * is this block fully invalidated?
1488 */
1489 if (offset <= curr_off)
1490 discard_buffer(bh);
1491 curr_off = next_off;
1492 bh = next;
1493 } while (bh != head);
1494
1495 /*
1496 * We release buffers only if the entire page is being invalidated.
1497 * The get_block cached value has been unconditionally invalidated,
1498 * so real IO is not possible anymore.
1499 */
1500 if (length == PAGE_SIZE)
1501 try_to_release_page(page, 0);
1502out:
1503 return;
1504}
1505EXPORT_SYMBOL(block_invalidatepage);
1506
1507
1508/*
1509 * We attach and possibly dirty the buffers atomically wrt
1510 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1511 * is already excluded via the page lock.
1512 */
1513void create_empty_buffers(struct page *page,
1514 unsigned long blocksize, unsigned long b_state)
1515{
1516 struct buffer_head *bh, *head, *tail;
1517
1518 head = alloc_page_buffers(page, blocksize, true);
1519 bh = head;
1520 do {
1521 bh->b_state |= b_state;
1522 tail = bh;
1523 bh = bh->b_this_page;
1524 } while (bh);
1525 tail->b_this_page = head;
1526
1527 spin_lock(&page->mapping->private_lock);
1528 if (PageUptodate(page) || PageDirty(page)) {
1529 bh = head;
1530 do {
1531 if (PageDirty(page))
1532 set_buffer_dirty(bh);
1533 if (PageUptodate(page))
1534 set_buffer_uptodate(bh);
1535 bh = bh->b_this_page;
1536 } while (bh != head);
1537 }
1538 attach_page_buffers(page, head);
1539 spin_unlock(&page->mapping->private_lock);
1540}
1541EXPORT_SYMBOL(create_empty_buffers);
1542
1543/**
1544 * clean_bdev_aliases: clean a range of buffers in block device
1545 * @bdev: Block device to clean buffers in
1546 * @block: Start of a range of blocks to clean
1547 * @len: Number of blocks to clean
1548 *
1549 * We are taking a range of blocks for data and we don't want writeback of any
1550 * buffer-cache aliases starting from return from this function and until the
1551 * moment when something will explicitly mark the buffer dirty (hopefully that
1552 * will not happen until we will free that block ;-) We don't even need to mark
1553 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1554 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1555 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1556 * would confuse anyone who might pick it with bread() afterwards...
1557 *
1558 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1559 * writeout I/O going on against recently-freed buffers. We don't wait on that
1560 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1561 * need to. That happens here.
1562 */
1563void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1564{
1565 struct inode *bd_inode = bdev->bd_inode;
1566 struct address_space *bd_mapping = bd_inode->i_mapping;
1567 struct pagevec pvec;
1568 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1569 pgoff_t end;
1570 int i, count;
1571 struct buffer_head *bh;
1572 struct buffer_head *head;
1573
1574 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1575 pagevec_init(&pvec);
1576 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1577 count = pagevec_count(&pvec);
1578 for (i = 0; i < count; i++) {
1579 struct page *page = pvec.pages[i];
1580
1581 if (!page_has_buffers(page))
1582 continue;
1583 /*
1584 * We use page lock instead of bd_mapping->private_lock
1585 * to pin buffers here since we can afford to sleep and
1586 * it scales better than a global spinlock lock.
1587 */
1588 lock_page(page);
1589 /* Recheck when the page is locked which pins bhs */
1590 if (!page_has_buffers(page))
1591 goto unlock_page;
1592 head = page_buffers(page);
1593 bh = head;
1594 do {
1595 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1596 goto next;
1597 if (bh->b_blocknr >= block + len)
1598 break;
1599 clear_buffer_dirty(bh);
1600 wait_on_buffer(bh);
1601 clear_buffer_req(bh);
1602next:
1603 bh = bh->b_this_page;
1604 } while (bh != head);
1605unlock_page:
1606 unlock_page(page);
1607 }
1608 pagevec_release(&pvec);
1609 cond_resched();
1610 /* End of range already reached? */
1611 if (index > end || !index)
1612 break;
1613 }
1614}
1615EXPORT_SYMBOL(clean_bdev_aliases);
1616
1617/*
1618 * Size is a power-of-two in the range 512..PAGE_SIZE,
1619 * and the case we care about most is PAGE_SIZE.
1620 *
1621 * So this *could* possibly be written with those
1622 * constraints in mind (relevant mostly if some
1623 * architecture has a slow bit-scan instruction)
1624 */
1625static inline int block_size_bits(unsigned int blocksize)
1626{
1627 return ilog2(blocksize);
1628}
1629
1630static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1631{
1632 BUG_ON(!PageLocked(page));
1633
1634 if (!page_has_buffers(page))
1635 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1636 b_state);
1637 return page_buffers(page);
1638}
1639
1640/*
1641 * NOTE! All mapped/uptodate combinations are valid:
1642 *
1643 * Mapped Uptodate Meaning
1644 *
1645 * No No "unknown" - must do get_block()
1646 * No Yes "hole" - zero-filled
1647 * Yes No "allocated" - allocated on disk, not read in
1648 * Yes Yes "valid" - allocated and up-to-date in memory.
1649 *
1650 * "Dirty" is valid only with the last case (mapped+uptodate).
1651 */
1652
1653/*
1654 * While block_write_full_page is writing back the dirty buffers under
1655 * the page lock, whoever dirtied the buffers may decide to clean them
1656 * again at any time. We handle that by only looking at the buffer
1657 * state inside lock_buffer().
1658 *
1659 * If block_write_full_page() is called for regular writeback
1660 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1661 * locked buffer. This only can happen if someone has written the buffer
1662 * directly, with submit_bh(). At the address_space level PageWriteback
1663 * prevents this contention from occurring.
1664 *
1665 * If block_write_full_page() is called with wbc->sync_mode ==
1666 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1667 * causes the writes to be flagged as synchronous writes.
1668 */
1669int __block_write_full_page(struct inode *inode, struct page *page,
1670 get_block_t *get_block, struct writeback_control *wbc,
1671 bh_end_io_t *handler)
1672{
1673 int err;
1674 sector_t block;
1675 sector_t last_block;
1676 struct buffer_head *bh, *head;
1677 unsigned int blocksize, bbits;
1678 int nr_underway = 0;
1679 int write_flags = wbc_to_write_flags(wbc);
1680
1681 head = create_page_buffers(page, inode,
1682 (1 << BH_Dirty)|(1 << BH_Uptodate));
1683
1684 /*
1685 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1686 * here, and the (potentially unmapped) buffers may become dirty at
1687 * any time. If a buffer becomes dirty here after we've inspected it
1688 * then we just miss that fact, and the page stays dirty.
1689 *
1690 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1691 * handle that here by just cleaning them.
1692 */
1693
1694 bh = head;
1695 blocksize = bh->b_size;
1696 bbits = block_size_bits(blocksize);
1697
1698 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1699 last_block = (i_size_read(inode) - 1) >> bbits;
1700
1701 /*
1702 * Get all the dirty buffers mapped to disk addresses and
1703 * handle any aliases from the underlying blockdev's mapping.
1704 */
1705 do {
1706 if (block > last_block) {
1707 /*
1708 * mapped buffers outside i_size will occur, because
1709 * this page can be outside i_size when there is a
1710 * truncate in progress.
1711 */
1712 /*
1713 * The buffer was zeroed by block_write_full_page()
1714 */
1715 clear_buffer_dirty(bh);
1716 set_buffer_uptodate(bh);
1717 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1718 buffer_dirty(bh)) {
1719 WARN_ON(bh->b_size != blocksize);
1720 err = get_block(inode, block, bh, 1);
1721 if (err)
1722 goto recover;
1723 clear_buffer_delay(bh);
1724 if (buffer_new(bh)) {
1725 /* blockdev mappings never come here */
1726 clear_buffer_new(bh);
1727 clean_bdev_bh_alias(bh);
1728 }
1729 }
1730 bh = bh->b_this_page;
1731 block++;
1732 } while (bh != head);
1733
1734 do {
1735 if (!buffer_mapped(bh))
1736 continue;
1737 /*
1738 * If it's a fully non-blocking write attempt and we cannot
1739 * lock the buffer then redirty the page. Note that this can
1740 * potentially cause a busy-wait loop from writeback threads
1741 * and kswapd activity, but those code paths have their own
1742 * higher-level throttling.
1743 */
1744 if (wbc->sync_mode != WB_SYNC_NONE) {
1745 lock_buffer(bh);
1746 } else if (!trylock_buffer(bh)) {
1747 redirty_page_for_writepage(wbc, page);
1748 continue;
1749 }
1750 if (test_clear_buffer_dirty(bh)) {
1751 mark_buffer_async_write_endio(bh, handler);
1752 } else {
1753 unlock_buffer(bh);
1754 }
1755 } while ((bh = bh->b_this_page) != head);
1756
1757 /*
1758 * The page and its buffers are protected by PageWriteback(), so we can
1759 * drop the bh refcounts early.
1760 */
1761 BUG_ON(PageWriteback(page));
1762 set_page_writeback(page);
1763
1764 do {
1765 struct buffer_head *next = bh->b_this_page;
1766 if (buffer_async_write(bh)) {
1767 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1768 inode->i_write_hint, wbc);
1769 nr_underway++;
1770 }
1771 bh = next;
1772 } while (bh != head);
1773 unlock_page(page);
1774
1775 err = 0;
1776done:
1777 if (nr_underway == 0) {
1778 /*
1779 * The page was marked dirty, but the buffers were
1780 * clean. Someone wrote them back by hand with
1781 * ll_rw_block/submit_bh. A rare case.
1782 */
1783 end_page_writeback(page);
1784
1785 /*
1786 * The page and buffer_heads can be released at any time from
1787 * here on.
1788 */
1789 }
1790 return err;
1791
1792recover:
1793 /*
1794 * ENOSPC, or some other error. We may already have added some
1795 * blocks to the file, so we need to write these out to avoid
1796 * exposing stale data.
1797 * The page is currently locked and not marked for writeback
1798 */
1799 bh = head;
1800 /* Recovery: lock and submit the mapped buffers */
1801 do {
1802 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1803 !buffer_delay(bh)) {
1804 lock_buffer(bh);
1805 mark_buffer_async_write_endio(bh, handler);
1806 } else {
1807 /*
1808 * The buffer may have been set dirty during
1809 * attachment to a dirty page.
1810 */
1811 clear_buffer_dirty(bh);
1812 }
1813 } while ((bh = bh->b_this_page) != head);
1814 SetPageError(page);
1815 BUG_ON(PageWriteback(page));
1816 mapping_set_error(page->mapping, err);
1817 set_page_writeback(page);
1818 do {
1819 struct buffer_head *next = bh->b_this_page;
1820 if (buffer_async_write(bh)) {
1821 clear_buffer_dirty(bh);
1822 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1823 inode->i_write_hint, wbc);
1824 nr_underway++;
1825 }
1826 bh = next;
1827 } while (bh != head);
1828 unlock_page(page);
1829 goto done;
1830}
1831EXPORT_SYMBOL(__block_write_full_page);
1832
1833/*
1834 * If a page has any new buffers, zero them out here, and mark them uptodate
1835 * and dirty so they'll be written out (in order to prevent uninitialised
1836 * block data from leaking). And clear the new bit.
1837 */
1838void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1839{
1840 unsigned int block_start, block_end;
1841 struct buffer_head *head, *bh;
1842
1843 BUG_ON(!PageLocked(page));
1844 if (!page_has_buffers(page))
1845 return;
1846
1847 bh = head = page_buffers(page);
1848 block_start = 0;
1849 do {
1850 block_end = block_start + bh->b_size;
1851
1852 if (buffer_new(bh)) {
1853 if (block_end > from && block_start < to) {
1854 if (!PageUptodate(page)) {
1855 unsigned start, size;
1856
1857 start = max(from, block_start);
1858 size = min(to, block_end) - start;
1859
1860 zero_user(page, start, size);
1861 set_buffer_uptodate(bh);
1862 }
1863
1864 clear_buffer_new(bh);
1865 mark_buffer_dirty(bh);
1866 }
1867 }
1868
1869 block_start = block_end;
1870 bh = bh->b_this_page;
1871 } while (bh != head);
1872}
1873EXPORT_SYMBOL(page_zero_new_buffers);
1874
1875static void
1876iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1877 struct iomap *iomap)
1878{
1879 loff_t offset = block << inode->i_blkbits;
1880
1881 bh->b_bdev = iomap->bdev;
1882
1883 /*
1884 * Block points to offset in file we need to map, iomap contains
1885 * the offset at which the map starts. If the map ends before the
1886 * current block, then do not map the buffer and let the caller
1887 * handle it.
1888 */
1889 BUG_ON(offset >= iomap->offset + iomap->length);
1890
1891 switch (iomap->type) {
1892 case IOMAP_HOLE:
1893 /*
1894 * If the buffer is not up to date or beyond the current EOF,
1895 * we need to mark it as new to ensure sub-block zeroing is
1896 * executed if necessary.
1897 */
1898 if (!buffer_uptodate(bh) ||
1899 (offset >= i_size_read(inode)))
1900 set_buffer_new(bh);
1901 break;
1902 case IOMAP_DELALLOC:
1903 if (!buffer_uptodate(bh) ||
1904 (offset >= i_size_read(inode)))
1905 set_buffer_new(bh);
1906 set_buffer_uptodate(bh);
1907 set_buffer_mapped(bh);
1908 set_buffer_delay(bh);
1909 break;
1910 case IOMAP_UNWRITTEN:
1911 /*
1912 * For unwritten regions, we always need to ensure that regions
1913 * in the block we are not writing to are zeroed. Mark the
1914 * buffer as new to ensure this.
1915 */
1916 set_buffer_new(bh);
1917 set_buffer_unwritten(bh);
1918 /* FALLTHRU */
1919 case IOMAP_MAPPED:
1920 if ((iomap->flags & IOMAP_F_NEW) ||
1921 offset >= i_size_read(inode))
1922 set_buffer_new(bh);
1923 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1924 inode->i_blkbits;
1925 set_buffer_mapped(bh);
1926 break;
1927 }
1928}
1929
1930int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1931 get_block_t *get_block, struct iomap *iomap)
1932{
1933 unsigned from = pos & (PAGE_SIZE - 1);
1934 unsigned to = from + len;
1935 struct inode *inode = page->mapping->host;
1936 unsigned block_start, block_end;
1937 sector_t block;
1938 int err = 0;
1939 unsigned blocksize, bbits;
1940 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1941
1942 BUG_ON(!PageLocked(page));
1943 BUG_ON(from > PAGE_SIZE);
1944 BUG_ON(to > PAGE_SIZE);
1945 BUG_ON(from > to);
1946
1947 head = create_page_buffers(page, inode, 0);
1948 blocksize = head->b_size;
1949 bbits = block_size_bits(blocksize);
1950
1951 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1952
1953 for(bh = head, block_start = 0; bh != head || !block_start;
1954 block++, block_start=block_end, bh = bh->b_this_page) {
1955 block_end = block_start + blocksize;
1956 if (block_end <= from || block_start >= to) {
1957 if (PageUptodate(page)) {
1958 if (!buffer_uptodate(bh))
1959 set_buffer_uptodate(bh);
1960 }
1961 continue;
1962 }
1963 if (buffer_new(bh))
1964 clear_buffer_new(bh);
1965 if (!buffer_mapped(bh)) {
1966 WARN_ON(bh->b_size != blocksize);
1967 if (get_block) {
1968 err = get_block(inode, block, bh, 1);
1969 if (err)
1970 break;
1971 } else {
1972 iomap_to_bh(inode, block, bh, iomap);
1973 }
1974
1975 if (buffer_new(bh)) {
1976 clean_bdev_bh_alias(bh);
1977 if (PageUptodate(page)) {
1978 clear_buffer_new(bh);
1979 set_buffer_uptodate(bh);
1980 mark_buffer_dirty(bh);
1981 continue;
1982 }
1983 if (block_end > to || block_start < from)
1984 zero_user_segments(page,
1985 to, block_end,
1986 block_start, from);
1987 continue;
1988 }
1989 }
1990 if (PageUptodate(page)) {
1991 if (!buffer_uptodate(bh))
1992 set_buffer_uptodate(bh);
1993 continue;
1994 }
1995 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1996 !buffer_unwritten(bh) &&
1997 (block_start < from || block_end > to)) {
1998 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1999 *wait_bh++=bh;
2000 }
2001 }
2002 /*
2003 * If we issued read requests - let them complete.
2004 */
2005 while(wait_bh > wait) {
2006 wait_on_buffer(*--wait_bh);
2007 if (!buffer_uptodate(*wait_bh))
2008 err = -EIO;
2009 }
2010 if (unlikely(err))
2011 page_zero_new_buffers(page, from, to);
2012 return err;
2013}
2014
2015int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2016 get_block_t *get_block)
2017{
2018 return __block_write_begin_int(page, pos, len, get_block, NULL);
2019}
2020EXPORT_SYMBOL(__block_write_begin);
2021
2022static int __block_commit_write(struct inode *inode, struct page *page,
2023 unsigned from, unsigned to)
2024{
2025 unsigned block_start, block_end;
2026 int partial = 0;
2027 unsigned blocksize;
2028 struct buffer_head *bh, *head;
2029
2030 bh = head = page_buffers(page);
2031 blocksize = bh->b_size;
2032
2033 block_start = 0;
2034 do {
2035 block_end = block_start + blocksize;
2036 if (block_end <= from || block_start >= to) {
2037 if (!buffer_uptodate(bh))
2038 partial = 1;
2039 } else {
2040 set_buffer_uptodate(bh);
2041 mark_buffer_dirty(bh);
2042 }
2043 clear_buffer_new(bh);
2044
2045 block_start = block_end;
2046 bh = bh->b_this_page;
2047 } while (bh != head);
2048
2049 /*
2050 * If this is a partial write which happened to make all buffers
2051 * uptodate then we can optimize away a bogus readpage() for
2052 * the next read(). Here we 'discover' whether the page went
2053 * uptodate as a result of this (potentially partial) write.
2054 */
2055 if (!partial)
2056 SetPageUptodate(page);
2057 return 0;
2058}
2059
2060/*
2061 * block_write_begin takes care of the basic task of block allocation and
2062 * bringing partial write blocks uptodate first.
2063 *
2064 * The filesystem needs to handle block truncation upon failure.
2065 */
2066int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2067 unsigned flags, struct page **pagep, get_block_t *get_block)
2068{
2069 pgoff_t index = pos >> PAGE_SHIFT;
2070 struct page *page;
2071 int status;
2072
2073 page = grab_cache_page_write_begin(mapping, index, flags);
2074 if (!page)
2075 return -ENOMEM;
2076
2077 status = __block_write_begin(page, pos, len, get_block);
2078 if (unlikely(status)) {
2079 unlock_page(page);
2080 put_page(page);
2081 page = NULL;
2082 }
2083
2084 *pagep = page;
2085 return status;
2086}
2087EXPORT_SYMBOL(block_write_begin);
2088
2089int block_write_end(struct file *file, struct address_space *mapping,
2090 loff_t pos, unsigned len, unsigned copied,
2091 struct page *page, void *fsdata)
2092{
2093 struct inode *inode = mapping->host;
2094 unsigned start;
2095
2096 start = pos & (PAGE_SIZE - 1);
2097
2098 if (unlikely(copied < len)) {
2099 /*
2100 * The buffers that were written will now be uptodate, so we
2101 * don't have to worry about a readpage reading them and
2102 * overwriting a partial write. However if we have encountered
2103 * a short write and only partially written into a buffer, it
2104 * will not be marked uptodate, so a readpage might come in and
2105 * destroy our partial write.
2106 *
2107 * Do the simplest thing, and just treat any short write to a
2108 * non uptodate page as a zero-length write, and force the
2109 * caller to redo the whole thing.
2110 */
2111 if (!PageUptodate(page))
2112 copied = 0;
2113
2114 page_zero_new_buffers(page, start+copied, start+len);
2115 }
2116 flush_dcache_page(page);
2117
2118 /* This could be a short (even 0-length) commit */
2119 __block_commit_write(inode, page, start, start+copied);
2120
2121 return copied;
2122}
2123EXPORT_SYMBOL(block_write_end);
2124
2125int generic_write_end(struct file *file, struct address_space *mapping,
2126 loff_t pos, unsigned len, unsigned copied,
2127 struct page *page, void *fsdata)
2128{
2129 struct inode *inode = mapping->host;
2130 loff_t old_size = inode->i_size;
2131 bool i_size_changed = false;
2132
2133 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2134
2135 /*
2136 * No need to use i_size_read() here, the i_size cannot change under us
2137 * because we hold i_rwsem.
2138 *
2139 * But it's important to update i_size while still holding page lock:
2140 * page writeout could otherwise come in and zero beyond i_size.
2141 */
2142 if (pos + copied > inode->i_size) {
2143 i_size_write(inode, pos + copied);
2144 i_size_changed = true;
2145 }
2146
2147 unlock_page(page);
2148 put_page(page);
2149
2150 if (old_size < pos)
2151 pagecache_isize_extended(inode, old_size, pos);
2152 /*
2153 * Don't mark the inode dirty under page lock. First, it unnecessarily
2154 * makes the holding time of page lock longer. Second, it forces lock
2155 * ordering of page lock and transaction start for journaling
2156 * filesystems.
2157 */
2158 if (i_size_changed)
2159 mark_inode_dirty(inode);
2160 return copied;
2161}
2162EXPORT_SYMBOL(generic_write_end);
2163
2164/*
2165 * block_is_partially_uptodate checks whether buffers within a page are
2166 * uptodate or not.
2167 *
2168 * Returns true if all buffers which correspond to a file portion
2169 * we want to read are uptodate.
2170 */
2171int block_is_partially_uptodate(struct page *page, unsigned long from,
2172 unsigned long count)
2173{
2174 unsigned block_start, block_end, blocksize;
2175 unsigned to;
2176 struct buffer_head *bh, *head;
2177 int ret = 1;
2178
2179 if (!page_has_buffers(page))
2180 return 0;
2181
2182 head = page_buffers(page);
2183 blocksize = head->b_size;
2184 to = min_t(unsigned, PAGE_SIZE - from, count);
2185 to = from + to;
2186 if (from < blocksize && to > PAGE_SIZE - blocksize)
2187 return 0;
2188
2189 bh = head;
2190 block_start = 0;
2191 do {
2192 block_end = block_start + blocksize;
2193 if (block_end > from && block_start < to) {
2194 if (!buffer_uptodate(bh)) {
2195 ret = 0;
2196 break;
2197 }
2198 if (block_end >= to)
2199 break;
2200 }
2201 block_start = block_end;
2202 bh = bh->b_this_page;
2203 } while (bh != head);
2204
2205 return ret;
2206}
2207EXPORT_SYMBOL(block_is_partially_uptodate);
2208
2209/*
2210 * Generic "read page" function for block devices that have the normal
2211 * get_block functionality. This is most of the block device filesystems.
2212 * Reads the page asynchronously --- the unlock_buffer() and
2213 * set/clear_buffer_uptodate() functions propagate buffer state into the
2214 * page struct once IO has completed.
2215 */
2216int block_read_full_page(struct page *page, get_block_t *get_block)
2217{
2218 struct inode *inode = page->mapping->host;
2219 sector_t iblock, lblock;
2220 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2221 unsigned int blocksize, bbits;
2222 int nr, i;
2223 int fully_mapped = 1;
2224
2225 head = create_page_buffers(page, inode, 0);
2226 blocksize = head->b_size;
2227 bbits = block_size_bits(blocksize);
2228
2229 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2230 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2231 bh = head;
2232 nr = 0;
2233 i = 0;
2234
2235 do {
2236 if (buffer_uptodate(bh))
2237 continue;
2238
2239 if (!buffer_mapped(bh)) {
2240 int err = 0;
2241
2242 fully_mapped = 0;
2243 if (iblock < lblock) {
2244 WARN_ON(bh->b_size != blocksize);
2245 err = get_block(inode, iblock, bh, 0);
2246 if (err)
2247 SetPageError(page);
2248 }
2249 if (!buffer_mapped(bh)) {
2250 zero_user(page, i * blocksize, blocksize);
2251 if (!err)
2252 set_buffer_uptodate(bh);
2253 continue;
2254 }
2255 /*
2256 * get_block() might have updated the buffer
2257 * synchronously
2258 */
2259 if (buffer_uptodate(bh))
2260 continue;
2261 }
2262 arr[nr++] = bh;
2263 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2264
2265 if (fully_mapped)
2266 SetPageMappedToDisk(page);
2267
2268 if (!nr) {
2269 /*
2270 * All buffers are uptodate - we can set the page uptodate
2271 * as well. But not if get_block() returned an error.
2272 */
2273 if (!PageError(page))
2274 SetPageUptodate(page);
2275 unlock_page(page);
2276 return 0;
2277 }
2278
2279 /* Stage two: lock the buffers */
2280 for (i = 0; i < nr; i++) {
2281 bh = arr[i];
2282 lock_buffer(bh);
2283 mark_buffer_async_read(bh);
2284 }
2285
2286 /*
2287 * Stage 3: start the IO. Check for uptodateness
2288 * inside the buffer lock in case another process reading
2289 * the underlying blockdev brought it uptodate (the sct fix).
2290 */
2291 for (i = 0; i < nr; i++) {
2292 bh = arr[i];
2293 if (buffer_uptodate(bh))
2294 end_buffer_async_read(bh, 1);
2295 else
2296 submit_bh(REQ_OP_READ, 0, bh);
2297 }
2298 return 0;
2299}
2300EXPORT_SYMBOL(block_read_full_page);
2301
2302/* utility function for filesystems that need to do work on expanding
2303 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2304 * deal with the hole.
2305 */
2306int generic_cont_expand_simple(struct inode *inode, loff_t size)
2307{
2308 struct address_space *mapping = inode->i_mapping;
2309 struct page *page;
2310 void *fsdata;
2311 int err;
2312
2313 err = inode_newsize_ok(inode, size);
2314 if (err)
2315 goto out;
2316
2317 err = pagecache_write_begin(NULL, mapping, size, 0,
2318 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2319 if (err)
2320 goto out;
2321
2322 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2323 BUG_ON(err > 0);
2324
2325out:
2326 return err;
2327}
2328EXPORT_SYMBOL(generic_cont_expand_simple);
2329
2330static int cont_expand_zero(struct file *file, struct address_space *mapping,
2331 loff_t pos, loff_t *bytes)
2332{
2333 struct inode *inode = mapping->host;
2334 unsigned int blocksize = i_blocksize(inode);
2335 struct page *page;
2336 void *fsdata;
2337 pgoff_t index, curidx;
2338 loff_t curpos;
2339 unsigned zerofrom, offset, len;
2340 int err = 0;
2341
2342 index = pos >> PAGE_SHIFT;
2343 offset = pos & ~PAGE_MASK;
2344
2345 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2346 zerofrom = curpos & ~PAGE_MASK;
2347 if (zerofrom & (blocksize-1)) {
2348 *bytes |= (blocksize-1);
2349 (*bytes)++;
2350 }
2351 len = PAGE_SIZE - zerofrom;
2352
2353 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2354 &page, &fsdata);
2355 if (err)
2356 goto out;
2357 zero_user(page, zerofrom, len);
2358 err = pagecache_write_end(file, mapping, curpos, len, len,
2359 page, fsdata);
2360 if (err < 0)
2361 goto out;
2362 BUG_ON(err != len);
2363 err = 0;
2364
2365 balance_dirty_pages_ratelimited(mapping);
2366
2367 if (fatal_signal_pending(current)) {
2368 err = -EINTR;
2369 goto out;
2370 }
2371 }
2372
2373 /* page covers the boundary, find the boundary offset */
2374 if (index == curidx) {
2375 zerofrom = curpos & ~PAGE_MASK;
2376 /* if we will expand the thing last block will be filled */
2377 if (offset <= zerofrom) {
2378 goto out;
2379 }
2380 if (zerofrom & (blocksize-1)) {
2381 *bytes |= (blocksize-1);
2382 (*bytes)++;
2383 }
2384 len = offset - zerofrom;
2385
2386 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2387 &page, &fsdata);
2388 if (err)
2389 goto out;
2390 zero_user(page, zerofrom, len);
2391 err = pagecache_write_end(file, mapping, curpos, len, len,
2392 page, fsdata);
2393 if (err < 0)
2394 goto out;
2395 BUG_ON(err != len);
2396 err = 0;
2397 }
2398out:
2399 return err;
2400}
2401
2402/*
2403 * For moronic filesystems that do not allow holes in file.
2404 * We may have to extend the file.
2405 */
2406int cont_write_begin(struct file *file, struct address_space *mapping,
2407 loff_t pos, unsigned len, unsigned flags,
2408 struct page **pagep, void **fsdata,
2409 get_block_t *get_block, loff_t *bytes)
2410{
2411 struct inode *inode = mapping->host;
2412 unsigned int blocksize = i_blocksize(inode);
2413 unsigned int zerofrom;
2414 int err;
2415
2416 err = cont_expand_zero(file, mapping, pos, bytes);
2417 if (err)
2418 return err;
2419
2420 zerofrom = *bytes & ~PAGE_MASK;
2421 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2422 *bytes |= (blocksize-1);
2423 (*bytes)++;
2424 }
2425
2426 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2427}
2428EXPORT_SYMBOL(cont_write_begin);
2429
2430int block_commit_write(struct page *page, unsigned from, unsigned to)
2431{
2432 struct inode *inode = page->mapping->host;
2433 __block_commit_write(inode,page,from,to);
2434 return 0;
2435}
2436EXPORT_SYMBOL(block_commit_write);
2437
2438/*
2439 * block_page_mkwrite() is not allowed to change the file size as it gets
2440 * called from a page fault handler when a page is first dirtied. Hence we must
2441 * be careful to check for EOF conditions here. We set the page up correctly
2442 * for a written page which means we get ENOSPC checking when writing into
2443 * holes and correct delalloc and unwritten extent mapping on filesystems that
2444 * support these features.
2445 *
2446 * We are not allowed to take the i_mutex here so we have to play games to
2447 * protect against truncate races as the page could now be beyond EOF. Because
2448 * truncate writes the inode size before removing pages, once we have the
2449 * page lock we can determine safely if the page is beyond EOF. If it is not
2450 * beyond EOF, then the page is guaranteed safe against truncation until we
2451 * unlock the page.
2452 *
2453 * Direct callers of this function should protect against filesystem freezing
2454 * using sb_start_pagefault() - sb_end_pagefault() functions.
2455 */
2456int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2457 get_block_t get_block)
2458{
2459 struct page *page = vmf->page;
2460 struct inode *inode = file_inode(vma->vm_file);
2461 unsigned long end;
2462 loff_t size;
2463 int ret;
2464
2465 lock_page(page);
2466 size = i_size_read(inode);
2467 if ((page->mapping != inode->i_mapping) ||
2468 (page_offset(page) > size)) {
2469 /* We overload EFAULT to mean page got truncated */
2470 ret = -EFAULT;
2471 goto out_unlock;
2472 }
2473
2474 /* page is wholly or partially inside EOF */
2475 if (((page->index + 1) << PAGE_SHIFT) > size)
2476 end = size & ~PAGE_MASK;
2477 else
2478 end = PAGE_SIZE;
2479
2480 ret = __block_write_begin(page, 0, end, get_block);
2481 if (!ret)
2482 ret = block_commit_write(page, 0, end);
2483
2484 if (unlikely(ret < 0))
2485 goto out_unlock;
2486 set_page_dirty(page);
2487 wait_for_stable_page(page);
2488 return 0;
2489out_unlock:
2490 unlock_page(page);
2491 return ret;
2492}
2493EXPORT_SYMBOL(block_page_mkwrite);
2494
2495/*
2496 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2497 * immediately, while under the page lock. So it needs a special end_io
2498 * handler which does not touch the bh after unlocking it.
2499 */
2500static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2501{
2502 __end_buffer_read_notouch(bh, uptodate);
2503}
2504
2505/*
2506 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2507 * the page (converting it to circular linked list and taking care of page
2508 * dirty races).
2509 */
2510static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2511{
2512 struct buffer_head *bh;
2513
2514 BUG_ON(!PageLocked(page));
2515
2516 spin_lock(&page->mapping->private_lock);
2517 bh = head;
2518 do {
2519 if (PageDirty(page))
2520 set_buffer_dirty(bh);
2521 if (!bh->b_this_page)
2522 bh->b_this_page = head;
2523 bh = bh->b_this_page;
2524 } while (bh != head);
2525 attach_page_buffers(page, head);
2526 spin_unlock(&page->mapping->private_lock);
2527}
2528
2529/*
2530 * On entry, the page is fully not uptodate.
2531 * On exit the page is fully uptodate in the areas outside (from,to)
2532 * The filesystem needs to handle block truncation upon failure.
2533 */
2534int nobh_write_begin(struct address_space *mapping,
2535 loff_t pos, unsigned len, unsigned flags,
2536 struct page **pagep, void **fsdata,
2537 get_block_t *get_block)
2538{
2539 struct inode *inode = mapping->host;
2540 const unsigned blkbits = inode->i_blkbits;
2541 const unsigned blocksize = 1 << blkbits;
2542 struct buffer_head *head, *bh;
2543 struct page *page;
2544 pgoff_t index;
2545 unsigned from, to;
2546 unsigned block_in_page;
2547 unsigned block_start, block_end;
2548 sector_t block_in_file;
2549 int nr_reads = 0;
2550 int ret = 0;
2551 int is_mapped_to_disk = 1;
2552
2553 index = pos >> PAGE_SHIFT;
2554 from = pos & (PAGE_SIZE - 1);
2555 to = from + len;
2556
2557 page = grab_cache_page_write_begin(mapping, index, flags);
2558 if (!page)
2559 return -ENOMEM;
2560 *pagep = page;
2561 *fsdata = NULL;
2562
2563 if (page_has_buffers(page)) {
2564 ret = __block_write_begin(page, pos, len, get_block);
2565 if (unlikely(ret))
2566 goto out_release;
2567 return ret;
2568 }
2569
2570 if (PageMappedToDisk(page))
2571 return 0;
2572
2573 /*
2574 * Allocate buffers so that we can keep track of state, and potentially
2575 * attach them to the page if an error occurs. In the common case of
2576 * no error, they will just be freed again without ever being attached
2577 * to the page (which is all OK, because we're under the page lock).
2578 *
2579 * Be careful: the buffer linked list is a NULL terminated one, rather
2580 * than the circular one we're used to.
2581 */
2582 head = alloc_page_buffers(page, blocksize, false);
2583 if (!head) {
2584 ret = -ENOMEM;
2585 goto out_release;
2586 }
2587
2588 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2589
2590 /*
2591 * We loop across all blocks in the page, whether or not they are
2592 * part of the affected region. This is so we can discover if the
2593 * page is fully mapped-to-disk.
2594 */
2595 for (block_start = 0, block_in_page = 0, bh = head;
2596 block_start < PAGE_SIZE;
2597 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2598 int create;
2599
2600 block_end = block_start + blocksize;
2601 bh->b_state = 0;
2602 create = 1;
2603 if (block_start >= to)
2604 create = 0;
2605 ret = get_block(inode, block_in_file + block_in_page,
2606 bh, create);
2607 if (ret)
2608 goto failed;
2609 if (!buffer_mapped(bh))
2610 is_mapped_to_disk = 0;
2611 if (buffer_new(bh))
2612 clean_bdev_bh_alias(bh);
2613 if (PageUptodate(page)) {
2614 set_buffer_uptodate(bh);
2615 continue;
2616 }
2617 if (buffer_new(bh) || !buffer_mapped(bh)) {
2618 zero_user_segments(page, block_start, from,
2619 to, block_end);
2620 continue;
2621 }
2622 if (buffer_uptodate(bh))
2623 continue; /* reiserfs does this */
2624 if (block_start < from || block_end > to) {
2625 lock_buffer(bh);
2626 bh->b_end_io = end_buffer_read_nobh;
2627 submit_bh(REQ_OP_READ, 0, bh);
2628 nr_reads++;
2629 }
2630 }
2631
2632 if (nr_reads) {
2633 /*
2634 * The page is locked, so these buffers are protected from
2635 * any VM or truncate activity. Hence we don't need to care
2636 * for the buffer_head refcounts.
2637 */
2638 for (bh = head; bh; bh = bh->b_this_page) {
2639 wait_on_buffer(bh);
2640 if (!buffer_uptodate(bh))
2641 ret = -EIO;
2642 }
2643 if (ret)
2644 goto failed;
2645 }
2646
2647 if (is_mapped_to_disk)
2648 SetPageMappedToDisk(page);
2649
2650 *fsdata = head; /* to be released by nobh_write_end */
2651
2652 return 0;
2653
2654failed:
2655 BUG_ON(!ret);
2656 /*
2657 * Error recovery is a bit difficult. We need to zero out blocks that
2658 * were newly allocated, and dirty them to ensure they get written out.
2659 * Buffers need to be attached to the page at this point, otherwise
2660 * the handling of potential IO errors during writeout would be hard
2661 * (could try doing synchronous writeout, but what if that fails too?)
2662 */
2663 attach_nobh_buffers(page, head);
2664 page_zero_new_buffers(page, from, to);
2665
2666out_release:
2667 unlock_page(page);
2668 put_page(page);
2669 *pagep = NULL;
2670
2671 return ret;
2672}
2673EXPORT_SYMBOL(nobh_write_begin);
2674
2675int nobh_write_end(struct file *file, struct address_space *mapping,
2676 loff_t pos, unsigned len, unsigned copied,
2677 struct page *page, void *fsdata)
2678{
2679 struct inode *inode = page->mapping->host;
2680 struct buffer_head *head = fsdata;
2681 struct buffer_head *bh;
2682 BUG_ON(fsdata != NULL && page_has_buffers(page));
2683
2684 if (unlikely(copied < len) && head)
2685 attach_nobh_buffers(page, head);
2686 if (page_has_buffers(page))
2687 return generic_write_end(file, mapping, pos, len,
2688 copied, page, fsdata);
2689
2690 SetPageUptodate(page);
2691 set_page_dirty(page);
2692 if (pos+copied > inode->i_size) {
2693 i_size_write(inode, pos+copied);
2694 mark_inode_dirty(inode);
2695 }
2696
2697 unlock_page(page);
2698 put_page(page);
2699
2700 while (head) {
2701 bh = head;
2702 head = head->b_this_page;
2703 free_buffer_head(bh);
2704 }
2705
2706 return copied;
2707}
2708EXPORT_SYMBOL(nobh_write_end);
2709
2710/*
2711 * nobh_writepage() - based on block_full_write_page() except
2712 * that it tries to operate without attaching bufferheads to
2713 * the page.
2714 */
2715int nobh_writepage(struct page *page, get_block_t *get_block,
2716 struct writeback_control *wbc)
2717{
2718 struct inode * const inode = page->mapping->host;
2719 loff_t i_size = i_size_read(inode);
2720 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2721 unsigned offset;
2722 int ret;
2723
2724 /* Is the page fully inside i_size? */
2725 if (page->index < end_index)
2726 goto out;
2727
2728 /* Is the page fully outside i_size? (truncate in progress) */
2729 offset = i_size & (PAGE_SIZE-1);
2730 if (page->index >= end_index+1 || !offset) {
2731 /*
2732 * The page may have dirty, unmapped buffers. For example,
2733 * they may have been added in ext3_writepage(). Make them
2734 * freeable here, so the page does not leak.
2735 */
2736#if 0
2737 /* Not really sure about this - do we need this ? */
2738 if (page->mapping->a_ops->invalidatepage)
2739 page->mapping->a_ops->invalidatepage(page, offset);
2740#endif
2741 unlock_page(page);
2742 return 0; /* don't care */
2743 }
2744
2745 /*
2746 * The page straddles i_size. It must be zeroed out on each and every
2747 * writepage invocation because it may be mmapped. "A file is mapped
2748 * in multiples of the page size. For a file that is not a multiple of
2749 * the page size, the remaining memory is zeroed when mapped, and
2750 * writes to that region are not written out to the file."
2751 */
2752 zero_user_segment(page, offset, PAGE_SIZE);
2753out:
2754 ret = mpage_writepage(page, get_block, wbc);
2755 if (ret == -EAGAIN)
2756 ret = __block_write_full_page(inode, page, get_block, wbc,
2757 end_buffer_async_write);
2758 return ret;
2759}
2760EXPORT_SYMBOL(nobh_writepage);
2761
2762int nobh_truncate_page(struct address_space *mapping,
2763 loff_t from, get_block_t *get_block)
2764{
2765 pgoff_t index = from >> PAGE_SHIFT;
2766 unsigned offset = from & (PAGE_SIZE-1);
2767 unsigned blocksize;
2768 sector_t iblock;
2769 unsigned length, pos;
2770 struct inode *inode = mapping->host;
2771 struct page *page;
2772 struct buffer_head map_bh;
2773 int err;
2774
2775 blocksize = i_blocksize(inode);
2776 length = offset & (blocksize - 1);
2777
2778 /* Block boundary? Nothing to do */
2779 if (!length)
2780 return 0;
2781
2782 length = blocksize - length;
2783 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2784
2785 page = grab_cache_page(mapping, index);
2786 err = -ENOMEM;
2787 if (!page)
2788 goto out;
2789
2790 if (page_has_buffers(page)) {
2791has_buffers:
2792 unlock_page(page);
2793 put_page(page);
2794 return block_truncate_page(mapping, from, get_block);
2795 }
2796
2797 /* Find the buffer that contains "offset" */
2798 pos = blocksize;
2799 while (offset >= pos) {
2800 iblock++;
2801 pos += blocksize;
2802 }
2803
2804 map_bh.b_size = blocksize;
2805 map_bh.b_state = 0;
2806 err = get_block(inode, iblock, &map_bh, 0);
2807 if (err)
2808 goto unlock;
2809 /* unmapped? It's a hole - nothing to do */
2810 if (!buffer_mapped(&map_bh))
2811 goto unlock;
2812
2813 /* Ok, it's mapped. Make sure it's up-to-date */
2814 if (!PageUptodate(page)) {
2815 err = mapping->a_ops->readpage(NULL, page);
2816 if (err) {
2817 put_page(page);
2818 goto out;
2819 }
2820 lock_page(page);
2821 if (!PageUptodate(page)) {
2822 err = -EIO;
2823 goto unlock;
2824 }
2825 if (page_has_buffers(page))
2826 goto has_buffers;
2827 }
2828 zero_user(page, offset, length);
2829 set_page_dirty(page);
2830 err = 0;
2831
2832unlock:
2833 unlock_page(page);
2834 put_page(page);
2835out:
2836 return err;
2837}
2838EXPORT_SYMBOL(nobh_truncate_page);
2839
2840int block_truncate_page(struct address_space *mapping,
2841 loff_t from, get_block_t *get_block)
2842{
2843 pgoff_t index = from >> PAGE_SHIFT;
2844 unsigned offset = from & (PAGE_SIZE-1);
2845 unsigned blocksize;
2846 sector_t iblock;
2847 unsigned length, pos;
2848 struct inode *inode = mapping->host;
2849 struct page *page;
2850 struct buffer_head *bh;
2851 int err;
2852
2853 blocksize = i_blocksize(inode);
2854 length = offset & (blocksize - 1);
2855
2856 /* Block boundary? Nothing to do */
2857 if (!length)
2858 return 0;
2859
2860 length = blocksize - length;
2861 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2862
2863 page = grab_cache_page(mapping, index);
2864 err = -ENOMEM;
2865 if (!page)
2866 goto out;
2867
2868 if (!page_has_buffers(page))
2869 create_empty_buffers(page, blocksize, 0);
2870
2871 /* Find the buffer that contains "offset" */
2872 bh = page_buffers(page);
2873 pos = blocksize;
2874 while (offset >= pos) {
2875 bh = bh->b_this_page;
2876 iblock++;
2877 pos += blocksize;
2878 }
2879
2880 err = 0;
2881 if (!buffer_mapped(bh)) {
2882 WARN_ON(bh->b_size != blocksize);
2883 err = get_block(inode, iblock, bh, 0);
2884 if (err)
2885 goto unlock;
2886 /* unmapped? It's a hole - nothing to do */
2887 if (!buffer_mapped(bh))
2888 goto unlock;
2889 }
2890
2891 /* Ok, it's mapped. Make sure it's up-to-date */
2892 if (PageUptodate(page))
2893 set_buffer_uptodate(bh);
2894
2895 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2896 err = -EIO;
2897 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2898 wait_on_buffer(bh);
2899 /* Uhhuh. Read error. Complain and punt. */
2900 if (!buffer_uptodate(bh))
2901 goto unlock;
2902 }
2903
2904 zero_user(page, offset, length);
2905 mark_buffer_dirty(bh);
2906 err = 0;
2907
2908unlock:
2909 unlock_page(page);
2910 put_page(page);
2911out:
2912 return err;
2913}
2914EXPORT_SYMBOL(block_truncate_page);
2915
2916/*
2917 * The generic ->writepage function for buffer-backed address_spaces
2918 */
2919int block_write_full_page(struct page *page, get_block_t *get_block,
2920 struct writeback_control *wbc)
2921{
2922 struct inode * const inode = page->mapping->host;
2923 loff_t i_size = i_size_read(inode);
2924 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2925 unsigned offset;
2926
2927 /* Is the page fully inside i_size? */
2928 if (page->index < end_index)
2929 return __block_write_full_page(inode, page, get_block, wbc,
2930 end_buffer_async_write);
2931
2932 /* Is the page fully outside i_size? (truncate in progress) */
2933 offset = i_size & (PAGE_SIZE-1);
2934 if (page->index >= end_index+1 || !offset) {
2935 /*
2936 * The page may have dirty, unmapped buffers. For example,
2937 * they may have been added in ext3_writepage(). Make them
2938 * freeable here, so the page does not leak.
2939 */
2940 do_invalidatepage(page, 0, PAGE_SIZE);
2941 unlock_page(page);
2942 return 0; /* don't care */
2943 }
2944
2945 /*
2946 * The page straddles i_size. It must be zeroed out on each and every
2947 * writepage invocation because it may be mmapped. "A file is mapped
2948 * in multiples of the page size. For a file that is not a multiple of
2949 * the page size, the remaining memory is zeroed when mapped, and
2950 * writes to that region are not written out to the file."
2951 */
2952 zero_user_segment(page, offset, PAGE_SIZE);
2953 return __block_write_full_page(inode, page, get_block, wbc,
2954 end_buffer_async_write);
2955}
2956EXPORT_SYMBOL(block_write_full_page);
2957
2958sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2959 get_block_t *get_block)
2960{
2961 struct inode *inode = mapping->host;
2962 struct buffer_head tmp = {
2963 .b_size = i_blocksize(inode),
2964 };
2965
2966 get_block(inode, block, &tmp, 0);
2967 return tmp.b_blocknr;
2968}
2969EXPORT_SYMBOL(generic_block_bmap);
2970
2971static void end_bio_bh_io_sync(struct bio *bio)
2972{
2973 struct buffer_head *bh = bio->bi_private;
2974
2975 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2976 set_bit(BH_Quiet, &bh->b_state);
2977
2978 bh->b_end_io(bh, !bio->bi_status);
2979 bio_put(bio);
2980}
2981
2982/*
2983 * This allows us to do IO even on the odd last sectors
2984 * of a device, even if the block size is some multiple
2985 * of the physical sector size.
2986 *
2987 * We'll just truncate the bio to the size of the device,
2988 * and clear the end of the buffer head manually.
2989 *
2990 * Truly out-of-range accesses will turn into actual IO
2991 * errors, this only handles the "we need to be able to
2992 * do IO at the final sector" case.
2993 */
2994void guard_bio_eod(int op, struct bio *bio)
2995{
2996 sector_t maxsector;
2997 struct bio_vec *bvec = bio_last_bvec_all(bio);
2998 unsigned truncated_bytes;
2999 struct hd_struct *part;
3000
3001 rcu_read_lock();
3002 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
3003 if (part)
3004 maxsector = part_nr_sects_read(part);
3005 else
3006 maxsector = get_capacity(bio->bi_disk);
3007 rcu_read_unlock();
3008
3009 if (!maxsector)
3010 return;
3011
3012 /*
3013 * If the *whole* IO is past the end of the device,
3014 * let it through, and the IO layer will turn it into
3015 * an EIO.
3016 */
3017 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3018 return;
3019
3020 maxsector -= bio->bi_iter.bi_sector;
3021 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3022 return;
3023
3024 /* Uhhuh. We've got a bio that straddles the device size! */
3025 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3026
3027 /*
3028 * The bio contains more than one segment which spans EOD, just return
3029 * and let IO layer turn it into an EIO
3030 */
3031 if (truncated_bytes > bvec->bv_len)
3032 return;
3033
3034 /* Truncate the bio.. */
3035 bio->bi_iter.bi_size -= truncated_bytes;
3036 bvec->bv_len -= truncated_bytes;
3037
3038 /* ..and clear the end of the buffer for reads */
3039 if (op == REQ_OP_READ) {
3040 struct bio_vec bv;
3041
3042 mp_bvec_last_segment(bvec, &bv);
3043 zero_user(bv.bv_page, bv.bv_offset + bv.bv_len,
3044 truncated_bytes);
3045 }
3046}
3047
3048static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3049 enum rw_hint write_hint, struct writeback_control *wbc)
3050{
3051 struct bio *bio;
3052
3053 BUG_ON(!buffer_locked(bh));
3054 BUG_ON(!buffer_mapped(bh));
3055 BUG_ON(!bh->b_end_io);
3056 BUG_ON(buffer_delay(bh));
3057 BUG_ON(buffer_unwritten(bh));
3058
3059 /*
3060 * Only clear out a write error when rewriting
3061 */
3062 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3063 clear_buffer_write_io_error(bh);
3064
3065 /*
3066 * from here on down, it's all bio -- do the initial mapping,
3067 * submit_bio -> generic_make_request may further map this bio around
3068 */
3069 bio = bio_alloc(GFP_NOIO, 1);
3070
3071 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3072 bio_set_dev(bio, bh->b_bdev);
3073 bio->bi_write_hint = write_hint;
3074
3075 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3076 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3077
3078 bio->bi_end_io = end_bio_bh_io_sync;
3079 bio->bi_private = bh;
3080
3081 /* Take care of bh's that straddle the end of the device */
3082 guard_bio_eod(op, bio);
3083
3084 if (buffer_meta(bh))
3085 op_flags |= REQ_META;
3086 if (buffer_prio(bh))
3087 op_flags |= REQ_PRIO;
3088 bio_set_op_attrs(bio, op, op_flags);
3089
3090 if (wbc) {
3091 wbc_init_bio(wbc, bio);
3092 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3093 }
3094
3095 submit_bio(bio);
3096 return 0;
3097}
3098
3099int submit_bh(int op, int op_flags, struct buffer_head *bh)
3100{
3101 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3102}
3103EXPORT_SYMBOL(submit_bh);
3104
3105/**
3106 * ll_rw_block: low-level access to block devices (DEPRECATED)
3107 * @op: whether to %READ or %WRITE
3108 * @op_flags: req_flag_bits
3109 * @nr: number of &struct buffer_heads in the array
3110 * @bhs: array of pointers to &struct buffer_head
3111 *
3112 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3113 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3114 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3115 * %REQ_RAHEAD.
3116 *
3117 * This function drops any buffer that it cannot get a lock on (with the
3118 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3119 * request, and any buffer that appears to be up-to-date when doing read
3120 * request. Further it marks as clean buffers that are processed for
3121 * writing (the buffer cache won't assume that they are actually clean
3122 * until the buffer gets unlocked).
3123 *
3124 * ll_rw_block sets b_end_io to simple completion handler that marks
3125 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3126 * any waiters.
3127 *
3128 * All of the buffers must be for the same device, and must also be a
3129 * multiple of the current approved size for the device.
3130 */
3131void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3132{
3133 int i;
3134
3135 for (i = 0; i < nr; i++) {
3136 struct buffer_head *bh = bhs[i];
3137
3138 if (!trylock_buffer(bh))
3139 continue;
3140 if (op == WRITE) {
3141 if (test_clear_buffer_dirty(bh)) {
3142 bh->b_end_io = end_buffer_write_sync;
3143 get_bh(bh);
3144 submit_bh(op, op_flags, bh);
3145 continue;
3146 }
3147 } else {
3148 if (!buffer_uptodate(bh)) {
3149 bh->b_end_io = end_buffer_read_sync;
3150 get_bh(bh);
3151 submit_bh(op, op_flags, bh);
3152 continue;
3153 }
3154 }
3155 unlock_buffer(bh);
3156 }
3157}
3158EXPORT_SYMBOL(ll_rw_block);
3159
3160void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3161{
3162 lock_buffer(bh);
3163 if (!test_clear_buffer_dirty(bh)) {
3164 unlock_buffer(bh);
3165 return;
3166 }
3167 bh->b_end_io = end_buffer_write_sync;
3168 get_bh(bh);
3169 submit_bh(REQ_OP_WRITE, op_flags, bh);
3170}
3171EXPORT_SYMBOL(write_dirty_buffer);
3172
3173/*
3174 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3175 * and then start new I/O and then wait upon it. The caller must have a ref on
3176 * the buffer_head.
3177 */
3178int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3179{
3180 int ret = 0;
3181
3182 WARN_ON(atomic_read(&bh->b_count) < 1);
3183 lock_buffer(bh);
3184 if (test_clear_buffer_dirty(bh)) {
3185 get_bh(bh);
3186 bh->b_end_io = end_buffer_write_sync;
3187 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3188 wait_on_buffer(bh);
3189 if (!ret && !buffer_uptodate(bh))
3190 ret = -EIO;
3191 } else {
3192 unlock_buffer(bh);
3193 }
3194 return ret;
3195}
3196EXPORT_SYMBOL(__sync_dirty_buffer);
3197
3198int sync_dirty_buffer(struct buffer_head *bh)
3199{
3200 return __sync_dirty_buffer(bh, REQ_SYNC);
3201}
3202EXPORT_SYMBOL(sync_dirty_buffer);
3203
3204/*
3205 * try_to_free_buffers() checks if all the buffers on this particular page
3206 * are unused, and releases them if so.
3207 *
3208 * Exclusion against try_to_free_buffers may be obtained by either
3209 * locking the page or by holding its mapping's private_lock.
3210 *
3211 * If the page is dirty but all the buffers are clean then we need to
3212 * be sure to mark the page clean as well. This is because the page
3213 * may be against a block device, and a later reattachment of buffers
3214 * to a dirty page will set *all* buffers dirty. Which would corrupt
3215 * filesystem data on the same device.
3216 *
3217 * The same applies to regular filesystem pages: if all the buffers are
3218 * clean then we set the page clean and proceed. To do that, we require
3219 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3220 * private_lock.
3221 *
3222 * try_to_free_buffers() is non-blocking.
3223 */
3224static inline int buffer_busy(struct buffer_head *bh)
3225{
3226 return atomic_read(&bh->b_count) |
3227 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3228}
3229
3230static int
3231drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3232{
3233 struct buffer_head *head = page_buffers(page);
3234 struct buffer_head *bh;
3235
3236 bh = head;
3237 do {
3238 if (buffer_busy(bh))
3239 goto failed;
3240 bh = bh->b_this_page;
3241 } while (bh != head);
3242
3243 do {
3244 struct buffer_head *next = bh->b_this_page;
3245
3246 if (bh->b_assoc_map)
3247 __remove_assoc_queue(bh);
3248 bh = next;
3249 } while (bh != head);
3250 *buffers_to_free = head;
3251 __clear_page_buffers(page);
3252 return 1;
3253failed:
3254 return 0;
3255}
3256
3257int try_to_free_buffers(struct page *page)
3258{
3259 struct address_space * const mapping = page->mapping;
3260 struct buffer_head *buffers_to_free = NULL;
3261 int ret = 0;
3262
3263 BUG_ON(!PageLocked(page));
3264 if (PageWriteback(page))
3265 return 0;
3266
3267 if (mapping == NULL) { /* can this still happen? */
3268 ret = drop_buffers(page, &buffers_to_free);
3269 goto out;
3270 }
3271
3272 spin_lock(&mapping->private_lock);
3273 ret = drop_buffers(page, &buffers_to_free);
3274
3275 /*
3276 * If the filesystem writes its buffers by hand (eg ext3)
3277 * then we can have clean buffers against a dirty page. We
3278 * clean the page here; otherwise the VM will never notice
3279 * that the filesystem did any IO at all.
3280 *
3281 * Also, during truncate, discard_buffer will have marked all
3282 * the page's buffers clean. We discover that here and clean
3283 * the page also.
3284 *
3285 * private_lock must be held over this entire operation in order
3286 * to synchronise against __set_page_dirty_buffers and prevent the
3287 * dirty bit from being lost.
3288 */
3289 if (ret)
3290 cancel_dirty_page(page);
3291 spin_unlock(&mapping->private_lock);
3292out:
3293 if (buffers_to_free) {
3294 struct buffer_head *bh = buffers_to_free;
3295
3296 do {
3297 struct buffer_head *next = bh->b_this_page;
3298 free_buffer_head(bh);
3299 bh = next;
3300 } while (bh != buffers_to_free);
3301 }
3302 return ret;
3303}
3304EXPORT_SYMBOL(try_to_free_buffers);
3305
3306/*
3307 * There are no bdflush tunables left. But distributions are
3308 * still running obsolete flush daemons, so we terminate them here.
3309 *
3310 * Use of bdflush() is deprecated and will be removed in a future kernel.
3311 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3312 */
3313SYSCALL_DEFINE2(bdflush, int, func, long, data)
3314{
3315 static int msg_count;
3316
3317 if (!capable(CAP_SYS_ADMIN))
3318 return -EPERM;
3319
3320 if (msg_count < 5) {
3321 msg_count++;
3322 printk(KERN_INFO
3323 "warning: process `%s' used the obsolete bdflush"
3324 " system call\n", current->comm);
3325 printk(KERN_INFO "Fix your initscripts?\n");
3326 }
3327
3328 if (func == 1)
3329 do_exit(0);
3330 return 0;
3331}
3332
3333/*
3334 * Buffer-head allocation
3335 */
3336static struct kmem_cache *bh_cachep __read_mostly;
3337
3338/*
3339 * Once the number of bh's in the machine exceeds this level, we start
3340 * stripping them in writeback.
3341 */
3342static unsigned long max_buffer_heads;
3343
3344int buffer_heads_over_limit;
3345
3346struct bh_accounting {
3347 int nr; /* Number of live bh's */
3348 int ratelimit; /* Limit cacheline bouncing */
3349};
3350
3351static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3352
3353static void recalc_bh_state(void)
3354{
3355 int i;
3356 int tot = 0;
3357
3358 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3359 return;
3360 __this_cpu_write(bh_accounting.ratelimit, 0);
3361 for_each_online_cpu(i)
3362 tot += per_cpu(bh_accounting, i).nr;
3363 buffer_heads_over_limit = (tot > max_buffer_heads);
3364}
3365
3366struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3367{
3368 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3369 if (ret) {
3370 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3371 preempt_disable();
3372 __this_cpu_inc(bh_accounting.nr);
3373 recalc_bh_state();
3374 preempt_enable();
3375 }
3376 return ret;
3377}
3378EXPORT_SYMBOL(alloc_buffer_head);
3379
3380void free_buffer_head(struct buffer_head *bh)
3381{
3382 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3383 kmem_cache_free(bh_cachep, bh);
3384 preempt_disable();
3385 __this_cpu_dec(bh_accounting.nr);
3386 recalc_bh_state();
3387 preempt_enable();
3388}
3389EXPORT_SYMBOL(free_buffer_head);
3390
3391static int buffer_exit_cpu_dead(unsigned int cpu)
3392{
3393 int i;
3394 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3395
3396 for (i = 0; i < BH_LRU_SIZE; i++) {
3397 brelse(b->bhs[i]);
3398 b->bhs[i] = NULL;
3399 }
3400 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3401 per_cpu(bh_accounting, cpu).nr = 0;
3402 return 0;
3403}
3404
3405/**
3406 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3407 * @bh: struct buffer_head
3408 *
3409 * Return true if the buffer is up-to-date and false,
3410 * with the buffer locked, if not.
3411 */
3412int bh_uptodate_or_lock(struct buffer_head *bh)
3413{
3414 if (!buffer_uptodate(bh)) {
3415 lock_buffer(bh);
3416 if (!buffer_uptodate(bh))
3417 return 0;
3418 unlock_buffer(bh);
3419 }
3420 return 1;
3421}
3422EXPORT_SYMBOL(bh_uptodate_or_lock);
3423
3424/**
3425 * bh_submit_read - Submit a locked buffer for reading
3426 * @bh: struct buffer_head
3427 *
3428 * Returns zero on success and -EIO on error.
3429 */
3430int bh_submit_read(struct buffer_head *bh)
3431{
3432 BUG_ON(!buffer_locked(bh));
3433
3434 if (buffer_uptodate(bh)) {
3435 unlock_buffer(bh);
3436 return 0;
3437 }
3438
3439 get_bh(bh);
3440 bh->b_end_io = end_buffer_read_sync;
3441 submit_bh(REQ_OP_READ, 0, bh);
3442 wait_on_buffer(bh);
3443 if (buffer_uptodate(bh))
3444 return 0;
3445 return -EIO;
3446}
3447EXPORT_SYMBOL(bh_submit_read);
3448
3449void __init buffer_init(void)
3450{
3451 unsigned long nrpages;
3452 int ret;
3453
3454 bh_cachep = kmem_cache_create("buffer_head",
3455 sizeof(struct buffer_head), 0,
3456 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3457 SLAB_MEM_SPREAD),
3458 NULL);
3459
3460 /*
3461 * Limit the bh occupancy to 10% of ZONE_NORMAL
3462 */
3463 nrpages = (nr_free_buffer_pages() * 10) / 100;
3464 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3465 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3466 NULL, buffer_exit_cpu_dead);
3467 WARN_ON(ret < 0);
3468}