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/*
2 * linux/fs/buffer.c
3 *
4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
5 */
6
7/*
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9 *
10 * Removed a lot of unnecessary code and simplified things now that
11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12 *
13 * Speed up hash, lru, and free list operations. Use gfp() for allocating
14 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
15 *
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17 *
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19 */
20
21#include <linux/kernel.h>
22#include <linux/sched/signal.h>
23#include <linux/syscalls.h>
24#include <linux/fs.h>
25#include <linux/iomap.h>
26#include <linux/mm.h>
27#include <linux/percpu.h>
28#include <linux/slab.h>
29#include <linux/capability.h>
30#include <linux/blkdev.h>
31#include <linux/file.h>
32#include <linux/quotaops.h>
33#include <linux/highmem.h>
34#include <linux/export.h>
35#include <linux/backing-dev.h>
36#include <linux/writeback.h>
37#include <linux/hash.h>
38#include <linux/suspend.h>
39#include <linux/buffer_head.h>
40#include <linux/task_io_accounting_ops.h>
41#include <linux/bio.h>
42#include <linux/notifier.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 <trace/events/block.h>
49
50static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
51static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
52 enum rw_hint hint, struct writeback_control *wbc);
53
54#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
55
56inline void touch_buffer(struct buffer_head *bh)
57{
58 trace_block_touch_buffer(bh);
59 mark_page_accessed(bh->b_page);
60}
61EXPORT_SYMBOL(touch_buffer);
62
63void __lock_buffer(struct buffer_head *bh)
64{
65 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
66}
67EXPORT_SYMBOL(__lock_buffer);
68
69void unlock_buffer(struct buffer_head *bh)
70{
71 clear_bit_unlock(BH_Lock, &bh->b_state);
72 smp_mb__after_atomic();
73 wake_up_bit(&bh->b_state, BH_Lock);
74}
75EXPORT_SYMBOL(unlock_buffer);
76
77/*
78 * Returns if the page has dirty or writeback buffers. If all the buffers
79 * are unlocked and clean then the PageDirty information is stale. If
80 * any of the pages are locked, it is assumed they are locked for IO.
81 */
82void buffer_check_dirty_writeback(struct page *page,
83 bool *dirty, bool *writeback)
84{
85 struct buffer_head *head, *bh;
86 *dirty = false;
87 *writeback = false;
88
89 BUG_ON(!PageLocked(page));
90
91 if (!page_has_buffers(page))
92 return;
93
94 if (PageWriteback(page))
95 *writeback = true;
96
97 head = page_buffers(page);
98 bh = head;
99 do {
100 if (buffer_locked(bh))
101 *writeback = true;
102
103 if (buffer_dirty(bh))
104 *dirty = true;
105
106 bh = bh->b_this_page;
107 } while (bh != head);
108}
109EXPORT_SYMBOL(buffer_check_dirty_writeback);
110
111/*
112 * Block until a buffer comes unlocked. This doesn't stop it
113 * from becoming locked again - you have to lock it yourself
114 * if you want to preserve its state.
115 */
116void __wait_on_buffer(struct buffer_head * bh)
117{
118 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
119}
120EXPORT_SYMBOL(__wait_on_buffer);
121
122static void
123__clear_page_buffers(struct page *page)
124{
125 ClearPagePrivate(page);
126 set_page_private(page, 0);
127 put_page(page);
128}
129
130static void buffer_io_error(struct buffer_head *bh, char *msg)
131{
132 if (!test_bit(BH_Quiet, &bh->b_state))
133 printk_ratelimited(KERN_ERR
134 "Buffer I/O error on dev %pg, logical block %llu%s\n",
135 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
136}
137
138/*
139 * End-of-IO handler helper function which does not touch the bh after
140 * unlocking it.
141 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
142 * a race there is benign: unlock_buffer() only use the bh's address for
143 * hashing after unlocking the buffer, so it doesn't actually touch the bh
144 * itself.
145 */
146static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
147{
148 if (uptodate) {
149 set_buffer_uptodate(bh);
150 } else {
151 /* This happens, due to failed read-ahead attempts. */
152 clear_buffer_uptodate(bh);
153 }
154 unlock_buffer(bh);
155}
156
157/*
158 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
159 * unlock the buffer. This is what ll_rw_block uses too.
160 */
161void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
162{
163 __end_buffer_read_notouch(bh, uptodate);
164 put_bh(bh);
165}
166EXPORT_SYMBOL(end_buffer_read_sync);
167
168void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
169{
170 if (uptodate) {
171 set_buffer_uptodate(bh);
172 } else {
173 buffer_io_error(bh, ", lost sync page write");
174 mark_buffer_write_io_error(bh);
175 clear_buffer_uptodate(bh);
176 }
177 unlock_buffer(bh);
178 put_bh(bh);
179}
180EXPORT_SYMBOL(end_buffer_write_sync);
181
182/*
183 * Various filesystems appear to want __find_get_block to be non-blocking.
184 * But it's the page lock which protects the buffers. To get around this,
185 * we get exclusion from try_to_free_buffers with the blockdev mapping's
186 * private_lock.
187 *
188 * Hack idea: for the blockdev mapping, private_lock contention
189 * may be quite high. This code could TryLock the page, and if that
190 * succeeds, there is no need to take private_lock.
191 */
192static struct buffer_head *
193__find_get_block_slow(struct block_device *bdev, sector_t block)
194{
195 struct inode *bd_inode = bdev->bd_inode;
196 struct address_space *bd_mapping = bd_inode->i_mapping;
197 struct buffer_head *ret = NULL;
198 pgoff_t index;
199 struct buffer_head *bh;
200 struct buffer_head *head;
201 struct page *page;
202 int all_mapped = 1;
203
204 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
205 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
206 if (!page)
207 goto out;
208
209 spin_lock(&bd_mapping->private_lock);
210 if (!page_has_buffers(page))
211 goto out_unlock;
212 head = page_buffers(page);
213 bh = head;
214 do {
215 if (!buffer_mapped(bh))
216 all_mapped = 0;
217 else if (bh->b_blocknr == block) {
218 ret = bh;
219 get_bh(bh);
220 goto out_unlock;
221 }
222 bh = bh->b_this_page;
223 } while (bh != head);
224
225 /* we might be here because some of the buffers on this page are
226 * not mapped. This is due to various races between
227 * file io on the block device and getblk. It gets dealt with
228 * elsewhere, don't buffer_error if we had some unmapped buffers
229 */
230 if (all_mapped) {
231 printk("__find_get_block_slow() failed. "
232 "block=%llu, b_blocknr=%llu\n",
233 (unsigned long long)block,
234 (unsigned long long)bh->b_blocknr);
235 printk("b_state=0x%08lx, b_size=%zu\n",
236 bh->b_state, bh->b_size);
237 printk("device %pg blocksize: %d\n", bdev,
238 1 << bd_inode->i_blkbits);
239 }
240out_unlock:
241 spin_unlock(&bd_mapping->private_lock);
242 put_page(page);
243out:
244 return ret;
245}
246
247/*
248 * I/O completion handler for block_read_full_page() - pages
249 * which come unlocked at the end of I/O.
250 */
251static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
252{
253 unsigned long flags;
254 struct buffer_head *first;
255 struct buffer_head *tmp;
256 struct page *page;
257 int page_uptodate = 1;
258
259 BUG_ON(!buffer_async_read(bh));
260
261 page = bh->b_page;
262 if (uptodate) {
263 set_buffer_uptodate(bh);
264 } else {
265 clear_buffer_uptodate(bh);
266 buffer_io_error(bh, ", async page read");
267 SetPageError(page);
268 }
269
270 /*
271 * Be _very_ careful from here on. Bad things can happen if
272 * two buffer heads end IO at almost the same time and both
273 * decide that the page is now completely done.
274 */
275 first = page_buffers(page);
276 local_irq_save(flags);
277 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
278 clear_buffer_async_read(bh);
279 unlock_buffer(bh);
280 tmp = bh;
281 do {
282 if (!buffer_uptodate(tmp))
283 page_uptodate = 0;
284 if (buffer_async_read(tmp)) {
285 BUG_ON(!buffer_locked(tmp));
286 goto still_busy;
287 }
288 tmp = tmp->b_this_page;
289 } while (tmp != bh);
290 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
291 local_irq_restore(flags);
292
293 /*
294 * If none of the buffers had errors and they are all
295 * uptodate then we can set the page uptodate.
296 */
297 if (page_uptodate && !PageError(page))
298 SetPageUptodate(page);
299 unlock_page(page);
300 return;
301
302still_busy:
303 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
304 local_irq_restore(flags);
305 return;
306}
307
308/*
309 * Completion handler for block_write_full_page() - pages which are unlocked
310 * during I/O, and which have PageWriteback cleared upon I/O completion.
311 */
312void end_buffer_async_write(struct buffer_head *bh, int uptodate)
313{
314 unsigned long flags;
315 struct buffer_head *first;
316 struct buffer_head *tmp;
317 struct page *page;
318
319 BUG_ON(!buffer_async_write(bh));
320
321 page = bh->b_page;
322 if (uptodate) {
323 set_buffer_uptodate(bh);
324 } else {
325 buffer_io_error(bh, ", lost async page write");
326 mark_buffer_write_io_error(bh);
327 clear_buffer_uptodate(bh);
328 SetPageError(page);
329 }
330
331 first = page_buffers(page);
332 local_irq_save(flags);
333 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
334
335 clear_buffer_async_write(bh);
336 unlock_buffer(bh);
337 tmp = bh->b_this_page;
338 while (tmp != bh) {
339 if (buffer_async_write(tmp)) {
340 BUG_ON(!buffer_locked(tmp));
341 goto still_busy;
342 }
343 tmp = tmp->b_this_page;
344 }
345 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
346 local_irq_restore(flags);
347 end_page_writeback(page);
348 return;
349
350still_busy:
351 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
352 local_irq_restore(flags);
353 return;
354}
355EXPORT_SYMBOL(end_buffer_async_write);
356
357/*
358 * If a page's buffers are under async readin (end_buffer_async_read
359 * completion) then there is a possibility that another thread of
360 * control could lock one of the buffers after it has completed
361 * but while some of the other buffers have not completed. This
362 * locked buffer would confuse end_buffer_async_read() into not unlocking
363 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
364 * that this buffer is not under async I/O.
365 *
366 * The page comes unlocked when it has no locked buffer_async buffers
367 * left.
368 *
369 * PageLocked prevents anyone starting new async I/O reads any of
370 * the buffers.
371 *
372 * PageWriteback is used to prevent simultaneous writeout of the same
373 * page.
374 *
375 * PageLocked prevents anyone from starting writeback of a page which is
376 * under read I/O (PageWriteback is only ever set against a locked page).
377 */
378static void mark_buffer_async_read(struct buffer_head *bh)
379{
380 bh->b_end_io = end_buffer_async_read;
381 set_buffer_async_read(bh);
382}
383
384static void mark_buffer_async_write_endio(struct buffer_head *bh,
385 bh_end_io_t *handler)
386{
387 bh->b_end_io = handler;
388 set_buffer_async_write(bh);
389}
390
391void mark_buffer_async_write(struct buffer_head *bh)
392{
393 mark_buffer_async_write_endio(bh, end_buffer_async_write);
394}
395EXPORT_SYMBOL(mark_buffer_async_write);
396
397
398/*
399 * fs/buffer.c contains helper functions for buffer-backed address space's
400 * fsync functions. A common requirement for buffer-based filesystems is
401 * that certain data from the backing blockdev needs to be written out for
402 * a successful fsync(). For example, ext2 indirect blocks need to be
403 * written back and waited upon before fsync() returns.
404 *
405 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
406 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
407 * management of a list of dependent buffers at ->i_mapping->private_list.
408 *
409 * Locking is a little subtle: try_to_free_buffers() will remove buffers
410 * from their controlling inode's queue when they are being freed. But
411 * try_to_free_buffers() will be operating against the *blockdev* mapping
412 * at the time, not against the S_ISREG file which depends on those buffers.
413 * So the locking for private_list is via the private_lock in the address_space
414 * which backs the buffers. Which is different from the address_space
415 * against which the buffers are listed. So for a particular address_space,
416 * mapping->private_lock does *not* protect mapping->private_list! In fact,
417 * mapping->private_list will always be protected by the backing blockdev's
418 * ->private_lock.
419 *
420 * Which introduces a requirement: all buffers on an address_space's
421 * ->private_list must be from the same address_space: the blockdev's.
422 *
423 * address_spaces which do not place buffers at ->private_list via these
424 * utility functions are free to use private_lock and private_list for
425 * whatever they want. The only requirement is that list_empty(private_list)
426 * be true at clear_inode() time.
427 *
428 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
429 * filesystems should do that. invalidate_inode_buffers() should just go
430 * BUG_ON(!list_empty).
431 *
432 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
433 * take an address_space, not an inode. And it should be called
434 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
435 * queued up.
436 *
437 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
438 * list if it is already on a list. Because if the buffer is on a list,
439 * it *must* already be on the right one. If not, the filesystem is being
440 * silly. This will save a ton of locking. But first we have to ensure
441 * that buffers are taken *off* the old inode's list when they are freed
442 * (presumably in truncate). That requires careful auditing of all
443 * filesystems (do it inside bforget()). It could also be done by bringing
444 * b_inode back.
445 */
446
447/*
448 * The buffer's backing address_space's private_lock must be held
449 */
450static void __remove_assoc_queue(struct buffer_head *bh)
451{
452 list_del_init(&bh->b_assoc_buffers);
453 WARN_ON(!bh->b_assoc_map);
454 bh->b_assoc_map = NULL;
455}
456
457int inode_has_buffers(struct inode *inode)
458{
459 return !list_empty(&inode->i_data.private_list);
460}
461
462/*
463 * osync is designed to support O_SYNC io. It waits synchronously for
464 * all already-submitted IO to complete, but does not queue any new
465 * writes to the disk.
466 *
467 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
468 * you dirty the buffers, and then use osync_inode_buffers to wait for
469 * completion. Any other dirty buffers which are not yet queued for
470 * write will not be flushed to disk by the osync.
471 */
472static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
473{
474 struct buffer_head *bh;
475 struct list_head *p;
476 int err = 0;
477
478 spin_lock(lock);
479repeat:
480 list_for_each_prev(p, list) {
481 bh = BH_ENTRY(p);
482 if (buffer_locked(bh)) {
483 get_bh(bh);
484 spin_unlock(lock);
485 wait_on_buffer(bh);
486 if (!buffer_uptodate(bh))
487 err = -EIO;
488 brelse(bh);
489 spin_lock(lock);
490 goto repeat;
491 }
492 }
493 spin_unlock(lock);
494 return err;
495}
496
497void emergency_thaw_bdev(struct super_block *sb)
498{
499 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
500 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
501}
502
503/**
504 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
505 * @mapping: the mapping which wants those buffers written
506 *
507 * Starts I/O against the buffers at mapping->private_list, and waits upon
508 * that I/O.
509 *
510 * Basically, this is a convenience function for fsync().
511 * @mapping is a file or directory which needs those buffers to be written for
512 * a successful fsync().
513 */
514int sync_mapping_buffers(struct address_space *mapping)
515{
516 struct address_space *buffer_mapping = mapping->private_data;
517
518 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
519 return 0;
520
521 return fsync_buffers_list(&buffer_mapping->private_lock,
522 &mapping->private_list);
523}
524EXPORT_SYMBOL(sync_mapping_buffers);
525
526/*
527 * Called when we've recently written block `bblock', and it is known that
528 * `bblock' was for a buffer_boundary() buffer. This means that the block at
529 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
530 * dirty, schedule it for IO. So that indirects merge nicely with their data.
531 */
532void write_boundary_block(struct block_device *bdev,
533 sector_t bblock, unsigned blocksize)
534{
535 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
536 if (bh) {
537 if (buffer_dirty(bh))
538 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
539 put_bh(bh);
540 }
541}
542
543void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
544{
545 struct address_space *mapping = inode->i_mapping;
546 struct address_space *buffer_mapping = bh->b_page->mapping;
547
548 mark_buffer_dirty(bh);
549 if (!mapping->private_data) {
550 mapping->private_data = buffer_mapping;
551 } else {
552 BUG_ON(mapping->private_data != buffer_mapping);
553 }
554 if (!bh->b_assoc_map) {
555 spin_lock(&buffer_mapping->private_lock);
556 list_move_tail(&bh->b_assoc_buffers,
557 &mapping->private_list);
558 bh->b_assoc_map = mapping;
559 spin_unlock(&buffer_mapping->private_lock);
560 }
561}
562EXPORT_SYMBOL(mark_buffer_dirty_inode);
563
564/*
565 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
566 * dirty.
567 *
568 * If warn is true, then emit a warning if the page is not uptodate and has
569 * not been truncated.
570 *
571 * The caller must hold lock_page_memcg().
572 */
573void __set_page_dirty(struct page *page, struct address_space *mapping,
574 int warn)
575{
576 unsigned long flags;
577
578 xa_lock_irqsave(&mapping->i_pages, flags);
579 if (page->mapping) { /* Race with truncate? */
580 WARN_ON_ONCE(warn && !PageUptodate(page));
581 account_page_dirtied(page, mapping);
582 radix_tree_tag_set(&mapping->i_pages,
583 page_index(page), PAGECACHE_TAG_DIRTY);
584 }
585 xa_unlock_irqrestore(&mapping->i_pages, flags);
586}
587EXPORT_SYMBOL_GPL(__set_page_dirty);
588
589/*
590 * Add a page to the dirty page list.
591 *
592 * It is a sad fact of life that this function is called from several places
593 * deeply under spinlocking. It may not sleep.
594 *
595 * If the page has buffers, the uptodate buffers are set dirty, to preserve
596 * dirty-state coherency between the page and the buffers. It the page does
597 * not have buffers then when they are later attached they will all be set
598 * dirty.
599 *
600 * The buffers are dirtied before the page is dirtied. There's a small race
601 * window in which a writepage caller may see the page cleanness but not the
602 * buffer dirtiness. That's fine. If this code were to set the page dirty
603 * before the buffers, a concurrent writepage caller could clear the page dirty
604 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
605 * page on the dirty page list.
606 *
607 * We use private_lock to lock against try_to_free_buffers while using the
608 * page's buffer list. Also use this to protect against clean buffers being
609 * added to the page after it was set dirty.
610 *
611 * FIXME: may need to call ->reservepage here as well. That's rather up to the
612 * address_space though.
613 */
614int __set_page_dirty_buffers(struct page *page)
615{
616 int newly_dirty;
617 struct address_space *mapping = page_mapping(page);
618
619 if (unlikely(!mapping))
620 return !TestSetPageDirty(page);
621
622 spin_lock(&mapping->private_lock);
623 if (page_has_buffers(page)) {
624 struct buffer_head *head = page_buffers(page);
625 struct buffer_head *bh = head;
626
627 do {
628 set_buffer_dirty(bh);
629 bh = bh->b_this_page;
630 } while (bh != head);
631 }
632 /*
633 * Lock out page->mem_cgroup migration to keep PageDirty
634 * synchronized with per-memcg dirty page counters.
635 */
636 lock_page_memcg(page);
637 newly_dirty = !TestSetPageDirty(page);
638 spin_unlock(&mapping->private_lock);
639
640 if (newly_dirty)
641 __set_page_dirty(page, mapping, 1);
642
643 unlock_page_memcg(page);
644
645 if (newly_dirty)
646 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
647
648 return newly_dirty;
649}
650EXPORT_SYMBOL(__set_page_dirty_buffers);
651
652/*
653 * Write out and wait upon a list of buffers.
654 *
655 * We have conflicting pressures: we want to make sure that all
656 * initially dirty buffers get waited on, but that any subsequently
657 * dirtied buffers don't. After all, we don't want fsync to last
658 * forever if somebody is actively writing to the file.
659 *
660 * Do this in two main stages: first we copy dirty buffers to a
661 * temporary inode list, queueing the writes as we go. Then we clean
662 * up, waiting for those writes to complete.
663 *
664 * During this second stage, any subsequent updates to the file may end
665 * up refiling the buffer on the original inode's dirty list again, so
666 * there is a chance we will end up with a buffer queued for write but
667 * not yet completed on that list. So, as a final cleanup we go through
668 * the osync code to catch these locked, dirty buffers without requeuing
669 * any newly dirty buffers for write.
670 */
671static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
672{
673 struct buffer_head *bh;
674 struct list_head tmp;
675 struct address_space *mapping;
676 int err = 0, err2;
677 struct blk_plug plug;
678
679 INIT_LIST_HEAD(&tmp);
680 blk_start_plug(&plug);
681
682 spin_lock(lock);
683 while (!list_empty(list)) {
684 bh = BH_ENTRY(list->next);
685 mapping = bh->b_assoc_map;
686 __remove_assoc_queue(bh);
687 /* Avoid race with mark_buffer_dirty_inode() which does
688 * a lockless check and we rely on seeing the dirty bit */
689 smp_mb();
690 if (buffer_dirty(bh) || buffer_locked(bh)) {
691 list_add(&bh->b_assoc_buffers, &tmp);
692 bh->b_assoc_map = mapping;
693 if (buffer_dirty(bh)) {
694 get_bh(bh);
695 spin_unlock(lock);
696 /*
697 * Ensure any pending I/O completes so that
698 * write_dirty_buffer() actually writes the
699 * current contents - it is a noop if I/O is
700 * still in flight on potentially older
701 * contents.
702 */
703 write_dirty_buffer(bh, REQ_SYNC);
704
705 /*
706 * Kick off IO for the previous mapping. Note
707 * that we will not run the very last mapping,
708 * wait_on_buffer() will do that for us
709 * through sync_buffer().
710 */
711 brelse(bh);
712 spin_lock(lock);
713 }
714 }
715 }
716
717 spin_unlock(lock);
718 blk_finish_plug(&plug);
719 spin_lock(lock);
720
721 while (!list_empty(&tmp)) {
722 bh = BH_ENTRY(tmp.prev);
723 get_bh(bh);
724 mapping = bh->b_assoc_map;
725 __remove_assoc_queue(bh);
726 /* Avoid race with mark_buffer_dirty_inode() which does
727 * a lockless check and we rely on seeing the dirty bit */
728 smp_mb();
729 if (buffer_dirty(bh)) {
730 list_add(&bh->b_assoc_buffers,
731 &mapping->private_list);
732 bh->b_assoc_map = mapping;
733 }
734 spin_unlock(lock);
735 wait_on_buffer(bh);
736 if (!buffer_uptodate(bh))
737 err = -EIO;
738 brelse(bh);
739 spin_lock(lock);
740 }
741
742 spin_unlock(lock);
743 err2 = osync_buffers_list(lock, list);
744 if (err)
745 return err;
746 else
747 return err2;
748}
749
750/*
751 * Invalidate any and all dirty buffers on a given inode. We are
752 * probably unmounting the fs, but that doesn't mean we have already
753 * done a sync(). Just drop the buffers from the inode list.
754 *
755 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
756 * assumes that all the buffers are against the blockdev. Not true
757 * for reiserfs.
758 */
759void invalidate_inode_buffers(struct inode *inode)
760{
761 if (inode_has_buffers(inode)) {
762 struct address_space *mapping = &inode->i_data;
763 struct list_head *list = &mapping->private_list;
764 struct address_space *buffer_mapping = mapping->private_data;
765
766 spin_lock(&buffer_mapping->private_lock);
767 while (!list_empty(list))
768 __remove_assoc_queue(BH_ENTRY(list->next));
769 spin_unlock(&buffer_mapping->private_lock);
770 }
771}
772EXPORT_SYMBOL(invalidate_inode_buffers);
773
774/*
775 * Remove any clean buffers from the inode's buffer list. This is called
776 * when we're trying to free the inode itself. Those buffers can pin it.
777 *
778 * Returns true if all buffers were removed.
779 */
780int remove_inode_buffers(struct inode *inode)
781{
782 int ret = 1;
783
784 if (inode_has_buffers(inode)) {
785 struct address_space *mapping = &inode->i_data;
786 struct list_head *list = &mapping->private_list;
787 struct address_space *buffer_mapping = mapping->private_data;
788
789 spin_lock(&buffer_mapping->private_lock);
790 while (!list_empty(list)) {
791 struct buffer_head *bh = BH_ENTRY(list->next);
792 if (buffer_dirty(bh)) {
793 ret = 0;
794 break;
795 }
796 __remove_assoc_queue(bh);
797 }
798 spin_unlock(&buffer_mapping->private_lock);
799 }
800 return ret;
801}
802
803/*
804 * Create the appropriate buffers when given a page for data area and
805 * the size of each buffer.. Use the bh->b_this_page linked list to
806 * follow the buffers created. Return NULL if unable to create more
807 * buffers.
808 *
809 * The retry flag is used to differentiate async IO (paging, swapping)
810 * which may not fail from ordinary buffer allocations.
811 */
812struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
813 bool retry)
814{
815 struct buffer_head *bh, *head;
816 gfp_t gfp = GFP_NOFS;
817 long offset;
818
819 if (retry)
820 gfp |= __GFP_NOFAIL;
821
822 head = NULL;
823 offset = PAGE_SIZE;
824 while ((offset -= size) >= 0) {
825 bh = alloc_buffer_head(gfp);
826 if (!bh)
827 goto no_grow;
828
829 bh->b_this_page = head;
830 bh->b_blocknr = -1;
831 head = bh;
832
833 bh->b_size = size;
834
835 /* Link the buffer to its page */
836 set_bh_page(bh, page, offset);
837 }
838 return head;
839/*
840 * In case anything failed, we just free everything we got.
841 */
842no_grow:
843 if (head) {
844 do {
845 bh = head;
846 head = head->b_this_page;
847 free_buffer_head(bh);
848 } while (head);
849 }
850
851 return NULL;
852}
853EXPORT_SYMBOL_GPL(alloc_page_buffers);
854
855static inline void
856link_dev_buffers(struct page *page, struct buffer_head *head)
857{
858 struct buffer_head *bh, *tail;
859
860 bh = head;
861 do {
862 tail = bh;
863 bh = bh->b_this_page;
864 } while (bh);
865 tail->b_this_page = head;
866 attach_page_buffers(page, head);
867}
868
869static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
870{
871 sector_t retval = ~((sector_t)0);
872 loff_t sz = i_size_read(bdev->bd_inode);
873
874 if (sz) {
875 unsigned int sizebits = blksize_bits(size);
876 retval = (sz >> sizebits);
877 }
878 return retval;
879}
880
881/*
882 * Initialise the state of a blockdev page's buffers.
883 */
884static sector_t
885init_page_buffers(struct page *page, struct block_device *bdev,
886 sector_t block, int size)
887{
888 struct buffer_head *head = page_buffers(page);
889 struct buffer_head *bh = head;
890 int uptodate = PageUptodate(page);
891 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
892
893 do {
894 if (!buffer_mapped(bh)) {
895 bh->b_end_io = NULL;
896 bh->b_private = NULL;
897 bh->b_bdev = bdev;
898 bh->b_blocknr = block;
899 if (uptodate)
900 set_buffer_uptodate(bh);
901 if (block < end_block)
902 set_buffer_mapped(bh);
903 }
904 block++;
905 bh = bh->b_this_page;
906 } while (bh != head);
907
908 /*
909 * Caller needs to validate requested block against end of device.
910 */
911 return end_block;
912}
913
914/*
915 * Create the page-cache page that contains the requested block.
916 *
917 * This is used purely for blockdev mappings.
918 */
919static int
920grow_dev_page(struct block_device *bdev, sector_t block,
921 pgoff_t index, int size, int sizebits, gfp_t gfp)
922{
923 struct inode *inode = bdev->bd_inode;
924 struct page *page;
925 struct buffer_head *bh;
926 sector_t end_block;
927 int ret = 0; /* Will call free_more_memory() */
928 gfp_t gfp_mask;
929
930 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
931
932 /*
933 * XXX: __getblk_slow() can not really deal with failure and
934 * will endlessly loop on improvised global reclaim. Prefer
935 * looping in the allocator rather than here, at least that
936 * code knows what it's doing.
937 */
938 gfp_mask |= __GFP_NOFAIL;
939
940 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
941
942 BUG_ON(!PageLocked(page));
943
944 if (page_has_buffers(page)) {
945 bh = page_buffers(page);
946 if (bh->b_size == size) {
947 end_block = init_page_buffers(page, bdev,
948 (sector_t)index << sizebits,
949 size);
950 goto done;
951 }
952 if (!try_to_free_buffers(page))
953 goto failed;
954 }
955
956 /*
957 * Allocate some buffers for this page
958 */
959 bh = alloc_page_buffers(page, size, true);
960
961 /*
962 * Link the page to the buffers and initialise them. Take the
963 * lock to be atomic wrt __find_get_block(), which does not
964 * run under the page lock.
965 */
966 spin_lock(&inode->i_mapping->private_lock);
967 link_dev_buffers(page, bh);
968 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
969 size);
970 spin_unlock(&inode->i_mapping->private_lock);
971done:
972 ret = (block < end_block) ? 1 : -ENXIO;
973failed:
974 unlock_page(page);
975 put_page(page);
976 return ret;
977}
978
979/*
980 * Create buffers for the specified block device block's page. If
981 * that page was dirty, the buffers are set dirty also.
982 */
983static int
984grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
985{
986 pgoff_t index;
987 int sizebits;
988
989 sizebits = -1;
990 do {
991 sizebits++;
992 } while ((size << sizebits) < PAGE_SIZE);
993
994 index = block >> sizebits;
995
996 /*
997 * Check for a block which wants to lie outside our maximum possible
998 * pagecache index. (this comparison is done using sector_t types).
999 */
1000 if (unlikely(index != block >> sizebits)) {
1001 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1002 "device %pg\n",
1003 __func__, (unsigned long long)block,
1004 bdev);
1005 return -EIO;
1006 }
1007
1008 /* Create a page with the proper size buffers.. */
1009 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1010}
1011
1012static struct buffer_head *
1013__getblk_slow(struct block_device *bdev, sector_t block,
1014 unsigned size, gfp_t gfp)
1015{
1016 /* Size must be multiple of hard sectorsize */
1017 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1018 (size < 512 || size > PAGE_SIZE))) {
1019 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1020 size);
1021 printk(KERN_ERR "logical block size: %d\n",
1022 bdev_logical_block_size(bdev));
1023
1024 dump_stack();
1025 return NULL;
1026 }
1027
1028 for (;;) {
1029 struct buffer_head *bh;
1030 int ret;
1031
1032 bh = __find_get_block(bdev, block, size);
1033 if (bh)
1034 return bh;
1035
1036 ret = grow_buffers(bdev, block, size, gfp);
1037 if (ret < 0)
1038 return NULL;
1039 }
1040}
1041
1042/*
1043 * The relationship between dirty buffers and dirty pages:
1044 *
1045 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1046 * the page is tagged dirty in its radix tree.
1047 *
1048 * At all times, the dirtiness of the buffers represents the dirtiness of
1049 * subsections of the page. If the page has buffers, the page dirty bit is
1050 * merely a hint about the true dirty state.
1051 *
1052 * When a page is set dirty in its entirety, all its buffers are marked dirty
1053 * (if the page has buffers).
1054 *
1055 * When a buffer is marked dirty, its page is dirtied, but the page's other
1056 * buffers are not.
1057 *
1058 * Also. When blockdev buffers are explicitly read with bread(), they
1059 * individually become uptodate. But their backing page remains not
1060 * uptodate - even if all of its buffers are uptodate. A subsequent
1061 * block_read_full_page() against that page will discover all the uptodate
1062 * buffers, will set the page uptodate and will perform no I/O.
1063 */
1064
1065/**
1066 * mark_buffer_dirty - mark a buffer_head as needing writeout
1067 * @bh: the buffer_head to mark dirty
1068 *
1069 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1070 * backing page dirty, then tag the page as dirty in its address_space's radix
1071 * tree and then attach the address_space's inode to its superblock's dirty
1072 * inode list.
1073 *
1074 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1075 * i_pages lock and mapping->host->i_lock.
1076 */
1077void mark_buffer_dirty(struct buffer_head *bh)
1078{
1079 WARN_ON_ONCE(!buffer_uptodate(bh));
1080
1081 trace_block_dirty_buffer(bh);
1082
1083 /*
1084 * Very *carefully* optimize the it-is-already-dirty case.
1085 *
1086 * Don't let the final "is it dirty" escape to before we
1087 * perhaps modified the buffer.
1088 */
1089 if (buffer_dirty(bh)) {
1090 smp_mb();
1091 if (buffer_dirty(bh))
1092 return;
1093 }
1094
1095 if (!test_set_buffer_dirty(bh)) {
1096 struct page *page = bh->b_page;
1097 struct address_space *mapping = NULL;
1098
1099 lock_page_memcg(page);
1100 if (!TestSetPageDirty(page)) {
1101 mapping = page_mapping(page);
1102 if (mapping)
1103 __set_page_dirty(page, mapping, 0);
1104 }
1105 unlock_page_memcg(page);
1106 if (mapping)
1107 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1108 }
1109}
1110EXPORT_SYMBOL(mark_buffer_dirty);
1111
1112void mark_buffer_write_io_error(struct buffer_head *bh)
1113{
1114 set_buffer_write_io_error(bh);
1115 /* FIXME: do we need to set this in both places? */
1116 if (bh->b_page && bh->b_page->mapping)
1117 mapping_set_error(bh->b_page->mapping, -EIO);
1118 if (bh->b_assoc_map)
1119 mapping_set_error(bh->b_assoc_map, -EIO);
1120}
1121EXPORT_SYMBOL(mark_buffer_write_io_error);
1122
1123/*
1124 * Decrement a buffer_head's reference count. If all buffers against a page
1125 * have zero reference count, are clean and unlocked, and if the page is clean
1126 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1127 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1128 * a page but it ends up not being freed, and buffers may later be reattached).
1129 */
1130void __brelse(struct buffer_head * buf)
1131{
1132 if (atomic_read(&buf->b_count)) {
1133 put_bh(buf);
1134 return;
1135 }
1136 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1137}
1138EXPORT_SYMBOL(__brelse);
1139
1140/*
1141 * bforget() is like brelse(), except it discards any
1142 * potentially dirty data.
1143 */
1144void __bforget(struct buffer_head *bh)
1145{
1146 clear_buffer_dirty(bh);
1147 if (bh->b_assoc_map) {
1148 struct address_space *buffer_mapping = bh->b_page->mapping;
1149
1150 spin_lock(&buffer_mapping->private_lock);
1151 list_del_init(&bh->b_assoc_buffers);
1152 bh->b_assoc_map = NULL;
1153 spin_unlock(&buffer_mapping->private_lock);
1154 }
1155 __brelse(bh);
1156}
1157EXPORT_SYMBOL(__bforget);
1158
1159static struct buffer_head *__bread_slow(struct buffer_head *bh)
1160{
1161 lock_buffer(bh);
1162 if (buffer_uptodate(bh)) {
1163 unlock_buffer(bh);
1164 return bh;
1165 } else {
1166 get_bh(bh);
1167 bh->b_end_io = end_buffer_read_sync;
1168 submit_bh(REQ_OP_READ, 0, bh);
1169 wait_on_buffer(bh);
1170 if (buffer_uptodate(bh))
1171 return bh;
1172 }
1173 brelse(bh);
1174 return NULL;
1175}
1176
1177/*
1178 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1179 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1180 * refcount elevated by one when they're in an LRU. A buffer can only appear
1181 * once in a particular CPU's LRU. A single buffer can be present in multiple
1182 * CPU's LRUs at the same time.
1183 *
1184 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1185 * sb_find_get_block().
1186 *
1187 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1188 * a local interrupt disable for that.
1189 */
1190
1191#define BH_LRU_SIZE 16
1192
1193struct bh_lru {
1194 struct buffer_head *bhs[BH_LRU_SIZE];
1195};
1196
1197static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1198
1199#ifdef CONFIG_SMP
1200#define bh_lru_lock() local_irq_disable()
1201#define bh_lru_unlock() local_irq_enable()
1202#else
1203#define bh_lru_lock() preempt_disable()
1204#define bh_lru_unlock() preempt_enable()
1205#endif
1206
1207static inline void check_irqs_on(void)
1208{
1209#ifdef irqs_disabled
1210 BUG_ON(irqs_disabled());
1211#endif
1212}
1213
1214/*
1215 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1216 * inserted at the front, and the buffer_head at the back if any is evicted.
1217 * Or, if already in the LRU it is moved to the front.
1218 */
1219static void bh_lru_install(struct buffer_head *bh)
1220{
1221 struct buffer_head *evictee = bh;
1222 struct bh_lru *b;
1223 int i;
1224
1225 check_irqs_on();
1226 bh_lru_lock();
1227
1228 b = this_cpu_ptr(&bh_lrus);
1229 for (i = 0; i < BH_LRU_SIZE; i++) {
1230 swap(evictee, b->bhs[i]);
1231 if (evictee == bh) {
1232 bh_lru_unlock();
1233 return;
1234 }
1235 }
1236
1237 get_bh(bh);
1238 bh_lru_unlock();
1239 brelse(evictee);
1240}
1241
1242/*
1243 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1244 */
1245static struct buffer_head *
1246lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1247{
1248 struct buffer_head *ret = NULL;
1249 unsigned int i;
1250
1251 check_irqs_on();
1252 bh_lru_lock();
1253 for (i = 0; i < BH_LRU_SIZE; i++) {
1254 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1255
1256 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1257 bh->b_size == size) {
1258 if (i) {
1259 while (i) {
1260 __this_cpu_write(bh_lrus.bhs[i],
1261 __this_cpu_read(bh_lrus.bhs[i - 1]));
1262 i--;
1263 }
1264 __this_cpu_write(bh_lrus.bhs[0], bh);
1265 }
1266 get_bh(bh);
1267 ret = bh;
1268 break;
1269 }
1270 }
1271 bh_lru_unlock();
1272 return ret;
1273}
1274
1275/*
1276 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1277 * it in the LRU and mark it as accessed. If it is not present then return
1278 * NULL
1279 */
1280struct buffer_head *
1281__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1282{
1283 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1284
1285 if (bh == NULL) {
1286 /* __find_get_block_slow will mark the page accessed */
1287 bh = __find_get_block_slow(bdev, block);
1288 if (bh)
1289 bh_lru_install(bh);
1290 } else
1291 touch_buffer(bh);
1292
1293 return bh;
1294}
1295EXPORT_SYMBOL(__find_get_block);
1296
1297/*
1298 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1299 * which corresponds to the passed block_device, block and size. The
1300 * returned buffer has its reference count incremented.
1301 *
1302 * __getblk_gfp() will lock up the machine if grow_dev_page's
1303 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1304 */
1305struct buffer_head *
1306__getblk_gfp(struct block_device *bdev, sector_t block,
1307 unsigned size, gfp_t gfp)
1308{
1309 struct buffer_head *bh = __find_get_block(bdev, block, size);
1310
1311 might_sleep();
1312 if (bh == NULL)
1313 bh = __getblk_slow(bdev, block, size, gfp);
1314 return bh;
1315}
1316EXPORT_SYMBOL(__getblk_gfp);
1317
1318/*
1319 * Do async read-ahead on a buffer..
1320 */
1321void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1322{
1323 struct buffer_head *bh = __getblk(bdev, block, size);
1324 if (likely(bh)) {
1325 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1326 brelse(bh);
1327 }
1328}
1329EXPORT_SYMBOL(__breadahead);
1330
1331/**
1332 * __bread_gfp() - reads a specified block and returns the bh
1333 * @bdev: the block_device to read from
1334 * @block: number of block
1335 * @size: size (in bytes) to read
1336 * @gfp: page allocation flag
1337 *
1338 * Reads a specified block, and returns buffer head that contains it.
1339 * The page cache can be allocated from non-movable area
1340 * not to prevent page migration if you set gfp to zero.
1341 * It returns NULL if the block was unreadable.
1342 */
1343struct buffer_head *
1344__bread_gfp(struct block_device *bdev, sector_t block,
1345 unsigned size, gfp_t gfp)
1346{
1347 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1348
1349 if (likely(bh) && !buffer_uptodate(bh))
1350 bh = __bread_slow(bh);
1351 return bh;
1352}
1353EXPORT_SYMBOL(__bread_gfp);
1354
1355/*
1356 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1357 * This doesn't race because it runs in each cpu either in irq
1358 * or with preempt disabled.
1359 */
1360static void invalidate_bh_lru(void *arg)
1361{
1362 struct bh_lru *b = &get_cpu_var(bh_lrus);
1363 int i;
1364
1365 for (i = 0; i < BH_LRU_SIZE; i++) {
1366 brelse(b->bhs[i]);
1367 b->bhs[i] = NULL;
1368 }
1369 put_cpu_var(bh_lrus);
1370}
1371
1372static bool has_bh_in_lru(int cpu, void *dummy)
1373{
1374 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1375 int i;
1376
1377 for (i = 0; i < BH_LRU_SIZE; i++) {
1378 if (b->bhs[i])
1379 return 1;
1380 }
1381
1382 return 0;
1383}
1384
1385void invalidate_bh_lrus(void)
1386{
1387 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1388}
1389EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1390
1391void set_bh_page(struct buffer_head *bh,
1392 struct page *page, unsigned long offset)
1393{
1394 bh->b_page = page;
1395 BUG_ON(offset >= PAGE_SIZE);
1396 if (PageHighMem(page))
1397 /*
1398 * This catches illegal uses and preserves the offset:
1399 */
1400 bh->b_data = (char *)(0 + offset);
1401 else
1402 bh->b_data = page_address(page) + offset;
1403}
1404EXPORT_SYMBOL(set_bh_page);
1405
1406/*
1407 * Called when truncating a buffer on a page completely.
1408 */
1409
1410/* Bits that are cleared during an invalidate */
1411#define BUFFER_FLAGS_DISCARD \
1412 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1413 1 << BH_Delay | 1 << BH_Unwritten)
1414
1415static void discard_buffer(struct buffer_head * bh)
1416{
1417 unsigned long b_state, b_state_old;
1418
1419 lock_buffer(bh);
1420 clear_buffer_dirty(bh);
1421 bh->b_bdev = NULL;
1422 b_state = bh->b_state;
1423 for (;;) {
1424 b_state_old = cmpxchg(&bh->b_state, b_state,
1425 (b_state & ~BUFFER_FLAGS_DISCARD));
1426 if (b_state_old == b_state)
1427 break;
1428 b_state = b_state_old;
1429 }
1430 unlock_buffer(bh);
1431}
1432
1433/**
1434 * block_invalidatepage - invalidate part or all of a buffer-backed page
1435 *
1436 * @page: the page which is affected
1437 * @offset: start of the range to invalidate
1438 * @length: length of the range to invalidate
1439 *
1440 * block_invalidatepage() is called when all or part of the page has become
1441 * invalidated by a truncate operation.
1442 *
1443 * block_invalidatepage() does not have to release all buffers, but it must
1444 * ensure that no dirty buffer is left outside @offset and that no I/O
1445 * is underway against any of the blocks which are outside the truncation
1446 * point. Because the caller is about to free (and possibly reuse) those
1447 * blocks on-disk.
1448 */
1449void block_invalidatepage(struct page *page, unsigned int offset,
1450 unsigned int length)
1451{
1452 struct buffer_head *head, *bh, *next;
1453 unsigned int curr_off = 0;
1454 unsigned int stop = length + offset;
1455
1456 BUG_ON(!PageLocked(page));
1457 if (!page_has_buffers(page))
1458 goto out;
1459
1460 /*
1461 * Check for overflow
1462 */
1463 BUG_ON(stop > PAGE_SIZE || stop < length);
1464
1465 head = page_buffers(page);
1466 bh = head;
1467 do {
1468 unsigned int next_off = curr_off + bh->b_size;
1469 next = bh->b_this_page;
1470
1471 /*
1472 * Are we still fully in range ?
1473 */
1474 if (next_off > stop)
1475 goto out;
1476
1477 /*
1478 * is this block fully invalidated?
1479 */
1480 if (offset <= curr_off)
1481 discard_buffer(bh);
1482 curr_off = next_off;
1483 bh = next;
1484 } while (bh != head);
1485
1486 /*
1487 * We release buffers only if the entire page is being invalidated.
1488 * The get_block cached value has been unconditionally invalidated,
1489 * so real IO is not possible anymore.
1490 */
1491 if (length == PAGE_SIZE)
1492 try_to_release_page(page, 0);
1493out:
1494 return;
1495}
1496EXPORT_SYMBOL(block_invalidatepage);
1497
1498
1499/*
1500 * We attach and possibly dirty the buffers atomically wrt
1501 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1502 * is already excluded via the page lock.
1503 */
1504void create_empty_buffers(struct page *page,
1505 unsigned long blocksize, unsigned long b_state)
1506{
1507 struct buffer_head *bh, *head, *tail;
1508
1509 head = alloc_page_buffers(page, blocksize, true);
1510 bh = head;
1511 do {
1512 bh->b_state |= b_state;
1513 tail = bh;
1514 bh = bh->b_this_page;
1515 } while (bh);
1516 tail->b_this_page = head;
1517
1518 spin_lock(&page->mapping->private_lock);
1519 if (PageUptodate(page) || PageDirty(page)) {
1520 bh = head;
1521 do {
1522 if (PageDirty(page))
1523 set_buffer_dirty(bh);
1524 if (PageUptodate(page))
1525 set_buffer_uptodate(bh);
1526 bh = bh->b_this_page;
1527 } while (bh != head);
1528 }
1529 attach_page_buffers(page, head);
1530 spin_unlock(&page->mapping->private_lock);
1531}
1532EXPORT_SYMBOL(create_empty_buffers);
1533
1534/**
1535 * clean_bdev_aliases: clean a range of buffers in block device
1536 * @bdev: Block device to clean buffers in
1537 * @block: Start of a range of blocks to clean
1538 * @len: Number of blocks to clean
1539 *
1540 * We are taking a range of blocks for data and we don't want writeback of any
1541 * buffer-cache aliases starting from return from this function and until the
1542 * moment when something will explicitly mark the buffer dirty (hopefully that
1543 * will not happen until we will free that block ;-) We don't even need to mark
1544 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1545 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1546 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1547 * would confuse anyone who might pick it with bread() afterwards...
1548 *
1549 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1550 * writeout I/O going on against recently-freed buffers. We don't wait on that
1551 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1552 * need to. That happens here.
1553 */
1554void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1555{
1556 struct inode *bd_inode = bdev->bd_inode;
1557 struct address_space *bd_mapping = bd_inode->i_mapping;
1558 struct pagevec pvec;
1559 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1560 pgoff_t end;
1561 int i, count;
1562 struct buffer_head *bh;
1563 struct buffer_head *head;
1564
1565 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1566 pagevec_init(&pvec);
1567 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1568 count = pagevec_count(&pvec);
1569 for (i = 0; i < count; i++) {
1570 struct page *page = pvec.pages[i];
1571
1572 if (!page_has_buffers(page))
1573 continue;
1574 /*
1575 * We use page lock instead of bd_mapping->private_lock
1576 * to pin buffers here since we can afford to sleep and
1577 * it scales better than a global spinlock lock.
1578 */
1579 lock_page(page);
1580 /* Recheck when the page is locked which pins bhs */
1581 if (!page_has_buffers(page))
1582 goto unlock_page;
1583 head = page_buffers(page);
1584 bh = head;
1585 do {
1586 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1587 goto next;
1588 if (bh->b_blocknr >= block + len)
1589 break;
1590 clear_buffer_dirty(bh);
1591 wait_on_buffer(bh);
1592 clear_buffer_req(bh);
1593next:
1594 bh = bh->b_this_page;
1595 } while (bh != head);
1596unlock_page:
1597 unlock_page(page);
1598 }
1599 pagevec_release(&pvec);
1600 cond_resched();
1601 /* End of range already reached? */
1602 if (index > end || !index)
1603 break;
1604 }
1605}
1606EXPORT_SYMBOL(clean_bdev_aliases);
1607
1608/*
1609 * Size is a power-of-two in the range 512..PAGE_SIZE,
1610 * and the case we care about most is PAGE_SIZE.
1611 *
1612 * So this *could* possibly be written with those
1613 * constraints in mind (relevant mostly if some
1614 * architecture has a slow bit-scan instruction)
1615 */
1616static inline int block_size_bits(unsigned int blocksize)
1617{
1618 return ilog2(blocksize);
1619}
1620
1621static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1622{
1623 BUG_ON(!PageLocked(page));
1624
1625 if (!page_has_buffers(page))
1626 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1627 b_state);
1628 return page_buffers(page);
1629}
1630
1631/*
1632 * NOTE! All mapped/uptodate combinations are valid:
1633 *
1634 * Mapped Uptodate Meaning
1635 *
1636 * No No "unknown" - must do get_block()
1637 * No Yes "hole" - zero-filled
1638 * Yes No "allocated" - allocated on disk, not read in
1639 * Yes Yes "valid" - allocated and up-to-date in memory.
1640 *
1641 * "Dirty" is valid only with the last case (mapped+uptodate).
1642 */
1643
1644/*
1645 * While block_write_full_page is writing back the dirty buffers under
1646 * the page lock, whoever dirtied the buffers may decide to clean them
1647 * again at any time. We handle that by only looking at the buffer
1648 * state inside lock_buffer().
1649 *
1650 * If block_write_full_page() is called for regular writeback
1651 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1652 * locked buffer. This only can happen if someone has written the buffer
1653 * directly, with submit_bh(). At the address_space level PageWriteback
1654 * prevents this contention from occurring.
1655 *
1656 * If block_write_full_page() is called with wbc->sync_mode ==
1657 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1658 * causes the writes to be flagged as synchronous writes.
1659 */
1660int __block_write_full_page(struct inode *inode, struct page *page,
1661 get_block_t *get_block, struct writeback_control *wbc,
1662 bh_end_io_t *handler)
1663{
1664 int err;
1665 sector_t block;
1666 sector_t last_block;
1667 struct buffer_head *bh, *head;
1668 unsigned int blocksize, bbits;
1669 int nr_underway = 0;
1670 int write_flags = wbc_to_write_flags(wbc);
1671
1672 head = create_page_buffers(page, inode,
1673 (1 << BH_Dirty)|(1 << BH_Uptodate));
1674
1675 /*
1676 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1677 * here, and the (potentially unmapped) buffers may become dirty at
1678 * any time. If a buffer becomes dirty here after we've inspected it
1679 * then we just miss that fact, and the page stays dirty.
1680 *
1681 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1682 * handle that here by just cleaning them.
1683 */
1684
1685 bh = head;
1686 blocksize = bh->b_size;
1687 bbits = block_size_bits(blocksize);
1688
1689 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1690 last_block = (i_size_read(inode) - 1) >> bbits;
1691
1692 /*
1693 * Get all the dirty buffers mapped to disk addresses and
1694 * handle any aliases from the underlying blockdev's mapping.
1695 */
1696 do {
1697 if (block > last_block) {
1698 /*
1699 * mapped buffers outside i_size will occur, because
1700 * this page can be outside i_size when there is a
1701 * truncate in progress.
1702 */
1703 /*
1704 * The buffer was zeroed by block_write_full_page()
1705 */
1706 clear_buffer_dirty(bh);
1707 set_buffer_uptodate(bh);
1708 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1709 buffer_dirty(bh)) {
1710 WARN_ON(bh->b_size != blocksize);
1711 err = get_block(inode, block, bh, 1);
1712 if (err)
1713 goto recover;
1714 clear_buffer_delay(bh);
1715 if (buffer_new(bh)) {
1716 /* blockdev mappings never come here */
1717 clear_buffer_new(bh);
1718 clean_bdev_bh_alias(bh);
1719 }
1720 }
1721 bh = bh->b_this_page;
1722 block++;
1723 } while (bh != head);
1724
1725 do {
1726 if (!buffer_mapped(bh))
1727 continue;
1728 /*
1729 * If it's a fully non-blocking write attempt and we cannot
1730 * lock the buffer then redirty the page. Note that this can
1731 * potentially cause a busy-wait loop from writeback threads
1732 * and kswapd activity, but those code paths have their own
1733 * higher-level throttling.
1734 */
1735 if (wbc->sync_mode != WB_SYNC_NONE) {
1736 lock_buffer(bh);
1737 } else if (!trylock_buffer(bh)) {
1738 redirty_page_for_writepage(wbc, page);
1739 continue;
1740 }
1741 if (test_clear_buffer_dirty(bh)) {
1742 mark_buffer_async_write_endio(bh, handler);
1743 } else {
1744 unlock_buffer(bh);
1745 }
1746 } while ((bh = bh->b_this_page) != head);
1747
1748 /*
1749 * The page and its buffers are protected by PageWriteback(), so we can
1750 * drop the bh refcounts early.
1751 */
1752 BUG_ON(PageWriteback(page));
1753 set_page_writeback(page);
1754
1755 do {
1756 struct buffer_head *next = bh->b_this_page;
1757 if (buffer_async_write(bh)) {
1758 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1759 inode->i_write_hint, wbc);
1760 nr_underway++;
1761 }
1762 bh = next;
1763 } while (bh != head);
1764 unlock_page(page);
1765
1766 err = 0;
1767done:
1768 if (nr_underway == 0) {
1769 /*
1770 * The page was marked dirty, but the buffers were
1771 * clean. Someone wrote them back by hand with
1772 * ll_rw_block/submit_bh. A rare case.
1773 */
1774 end_page_writeback(page);
1775
1776 /*
1777 * The page and buffer_heads can be released at any time from
1778 * here on.
1779 */
1780 }
1781 return err;
1782
1783recover:
1784 /*
1785 * ENOSPC, or some other error. We may already have added some
1786 * blocks to the file, so we need to write these out to avoid
1787 * exposing stale data.
1788 * The page is currently locked and not marked for writeback
1789 */
1790 bh = head;
1791 /* Recovery: lock and submit the mapped buffers */
1792 do {
1793 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1794 !buffer_delay(bh)) {
1795 lock_buffer(bh);
1796 mark_buffer_async_write_endio(bh, handler);
1797 } else {
1798 /*
1799 * The buffer may have been set dirty during
1800 * attachment to a dirty page.
1801 */
1802 clear_buffer_dirty(bh);
1803 }
1804 } while ((bh = bh->b_this_page) != head);
1805 SetPageError(page);
1806 BUG_ON(PageWriteback(page));
1807 mapping_set_error(page->mapping, err);
1808 set_page_writeback(page);
1809 do {
1810 struct buffer_head *next = bh->b_this_page;
1811 if (buffer_async_write(bh)) {
1812 clear_buffer_dirty(bh);
1813 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1814 inode->i_write_hint, wbc);
1815 nr_underway++;
1816 }
1817 bh = next;
1818 } while (bh != head);
1819 unlock_page(page);
1820 goto done;
1821}
1822EXPORT_SYMBOL(__block_write_full_page);
1823
1824/*
1825 * If a page has any new buffers, zero them out here, and mark them uptodate
1826 * and dirty so they'll be written out (in order to prevent uninitialised
1827 * block data from leaking). And clear the new bit.
1828 */
1829void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1830{
1831 unsigned int block_start, block_end;
1832 struct buffer_head *head, *bh;
1833
1834 BUG_ON(!PageLocked(page));
1835 if (!page_has_buffers(page))
1836 return;
1837
1838 bh = head = page_buffers(page);
1839 block_start = 0;
1840 do {
1841 block_end = block_start + bh->b_size;
1842
1843 if (buffer_new(bh)) {
1844 if (block_end > from && block_start < to) {
1845 if (!PageUptodate(page)) {
1846 unsigned start, size;
1847
1848 start = max(from, block_start);
1849 size = min(to, block_end) - start;
1850
1851 zero_user(page, start, size);
1852 set_buffer_uptodate(bh);
1853 }
1854
1855 clear_buffer_new(bh);
1856 mark_buffer_dirty(bh);
1857 }
1858 }
1859
1860 block_start = block_end;
1861 bh = bh->b_this_page;
1862 } while (bh != head);
1863}
1864EXPORT_SYMBOL(page_zero_new_buffers);
1865
1866static void
1867iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1868 struct iomap *iomap)
1869{
1870 loff_t offset = block << inode->i_blkbits;
1871
1872 bh->b_bdev = iomap->bdev;
1873
1874 /*
1875 * Block points to offset in file we need to map, iomap contains
1876 * the offset at which the map starts. If the map ends before the
1877 * current block, then do not map the buffer and let the caller
1878 * handle it.
1879 */
1880 BUG_ON(offset >= iomap->offset + iomap->length);
1881
1882 switch (iomap->type) {
1883 case IOMAP_HOLE:
1884 /*
1885 * If the buffer is not up to date or beyond the current EOF,
1886 * we need to mark it as new to ensure sub-block zeroing is
1887 * executed if necessary.
1888 */
1889 if (!buffer_uptodate(bh) ||
1890 (offset >= i_size_read(inode)))
1891 set_buffer_new(bh);
1892 break;
1893 case IOMAP_DELALLOC:
1894 if (!buffer_uptodate(bh) ||
1895 (offset >= i_size_read(inode)))
1896 set_buffer_new(bh);
1897 set_buffer_uptodate(bh);
1898 set_buffer_mapped(bh);
1899 set_buffer_delay(bh);
1900 break;
1901 case IOMAP_UNWRITTEN:
1902 /*
1903 * For unwritten regions, we always need to ensure that
1904 * sub-block writes cause the regions in the block we are not
1905 * writing to are zeroed. Set the buffer as new to ensure this.
1906 */
1907 set_buffer_new(bh);
1908 set_buffer_unwritten(bh);
1909 /* FALLTHRU */
1910 case IOMAP_MAPPED:
1911 if (offset >= i_size_read(inode))
1912 set_buffer_new(bh);
1913 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1914 inode->i_blkbits;
1915 set_buffer_mapped(bh);
1916 break;
1917 }
1918}
1919
1920int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1921 get_block_t *get_block, struct iomap *iomap)
1922{
1923 unsigned from = pos & (PAGE_SIZE - 1);
1924 unsigned to = from + len;
1925 struct inode *inode = page->mapping->host;
1926 unsigned block_start, block_end;
1927 sector_t block;
1928 int err = 0;
1929 unsigned blocksize, bbits;
1930 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1931
1932 BUG_ON(!PageLocked(page));
1933 BUG_ON(from > PAGE_SIZE);
1934 BUG_ON(to > PAGE_SIZE);
1935 BUG_ON(from > to);
1936
1937 head = create_page_buffers(page, inode, 0);
1938 blocksize = head->b_size;
1939 bbits = block_size_bits(blocksize);
1940
1941 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1942
1943 for(bh = head, block_start = 0; bh != head || !block_start;
1944 block++, block_start=block_end, bh = bh->b_this_page) {
1945 block_end = block_start + blocksize;
1946 if (block_end <= from || block_start >= to) {
1947 if (PageUptodate(page)) {
1948 if (!buffer_uptodate(bh))
1949 set_buffer_uptodate(bh);
1950 }
1951 continue;
1952 }
1953 if (buffer_new(bh))
1954 clear_buffer_new(bh);
1955 if (!buffer_mapped(bh)) {
1956 WARN_ON(bh->b_size != blocksize);
1957 if (get_block) {
1958 err = get_block(inode, block, bh, 1);
1959 if (err)
1960 break;
1961 } else {
1962 iomap_to_bh(inode, block, bh, iomap);
1963 }
1964
1965 if (buffer_new(bh)) {
1966 clean_bdev_bh_alias(bh);
1967 if (PageUptodate(page)) {
1968 clear_buffer_new(bh);
1969 set_buffer_uptodate(bh);
1970 mark_buffer_dirty(bh);
1971 continue;
1972 }
1973 if (block_end > to || block_start < from)
1974 zero_user_segments(page,
1975 to, block_end,
1976 block_start, from);
1977 continue;
1978 }
1979 }
1980 if (PageUptodate(page)) {
1981 if (!buffer_uptodate(bh))
1982 set_buffer_uptodate(bh);
1983 continue;
1984 }
1985 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1986 !buffer_unwritten(bh) &&
1987 (block_start < from || block_end > to)) {
1988 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1989 *wait_bh++=bh;
1990 }
1991 }
1992 /*
1993 * If we issued read requests - let them complete.
1994 */
1995 while(wait_bh > wait) {
1996 wait_on_buffer(*--wait_bh);
1997 if (!buffer_uptodate(*wait_bh))
1998 err = -EIO;
1999 }
2000 if (unlikely(err))
2001 page_zero_new_buffers(page, from, to);
2002 return err;
2003}
2004
2005int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2006 get_block_t *get_block)
2007{
2008 return __block_write_begin_int(page, pos, len, get_block, NULL);
2009}
2010EXPORT_SYMBOL(__block_write_begin);
2011
2012static int __block_commit_write(struct inode *inode, struct page *page,
2013 unsigned from, unsigned to)
2014{
2015 unsigned block_start, block_end;
2016 int partial = 0;
2017 unsigned blocksize;
2018 struct buffer_head *bh, *head;
2019
2020 bh = head = page_buffers(page);
2021 blocksize = bh->b_size;
2022
2023 block_start = 0;
2024 do {
2025 block_end = block_start + blocksize;
2026 if (block_end <= from || block_start >= to) {
2027 if (!buffer_uptodate(bh))
2028 partial = 1;
2029 } else {
2030 set_buffer_uptodate(bh);
2031 mark_buffer_dirty(bh);
2032 }
2033 clear_buffer_new(bh);
2034
2035 block_start = block_end;
2036 bh = bh->b_this_page;
2037 } while (bh != head);
2038
2039 /*
2040 * If this is a partial write which happened to make all buffers
2041 * uptodate then we can optimize away a bogus readpage() for
2042 * the next read(). Here we 'discover' whether the page went
2043 * uptodate as a result of this (potentially partial) write.
2044 */
2045 if (!partial)
2046 SetPageUptodate(page);
2047 return 0;
2048}
2049
2050/*
2051 * block_write_begin takes care of the basic task of block allocation and
2052 * bringing partial write blocks uptodate first.
2053 *
2054 * The filesystem needs to handle block truncation upon failure.
2055 */
2056int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2057 unsigned flags, struct page **pagep, get_block_t *get_block)
2058{
2059 pgoff_t index = pos >> PAGE_SHIFT;
2060 struct page *page;
2061 int status;
2062
2063 page = grab_cache_page_write_begin(mapping, index, flags);
2064 if (!page)
2065 return -ENOMEM;
2066
2067 status = __block_write_begin(page, pos, len, get_block);
2068 if (unlikely(status)) {
2069 unlock_page(page);
2070 put_page(page);
2071 page = NULL;
2072 }
2073
2074 *pagep = page;
2075 return status;
2076}
2077EXPORT_SYMBOL(block_write_begin);
2078
2079int block_write_end(struct file *file, struct address_space *mapping,
2080 loff_t pos, unsigned len, unsigned copied,
2081 struct page *page, void *fsdata)
2082{
2083 struct inode *inode = mapping->host;
2084 unsigned start;
2085
2086 start = pos & (PAGE_SIZE - 1);
2087
2088 if (unlikely(copied < len)) {
2089 /*
2090 * The buffers that were written will now be uptodate, so we
2091 * don't have to worry about a readpage reading them and
2092 * overwriting a partial write. However if we have encountered
2093 * a short write and only partially written into a buffer, it
2094 * will not be marked uptodate, so a readpage might come in and
2095 * destroy our partial write.
2096 *
2097 * Do the simplest thing, and just treat any short write to a
2098 * non uptodate page as a zero-length write, and force the
2099 * caller to redo the whole thing.
2100 */
2101 if (!PageUptodate(page))
2102 copied = 0;
2103
2104 page_zero_new_buffers(page, start+copied, start+len);
2105 }
2106 flush_dcache_page(page);
2107
2108 /* This could be a short (even 0-length) commit */
2109 __block_commit_write(inode, page, start, start+copied);
2110
2111 return copied;
2112}
2113EXPORT_SYMBOL(block_write_end);
2114
2115int generic_write_end(struct file *file, struct address_space *mapping,
2116 loff_t pos, unsigned len, unsigned copied,
2117 struct page *page, void *fsdata)
2118{
2119 struct inode *inode = mapping->host;
2120 loff_t old_size = inode->i_size;
2121 int i_size_changed = 0;
2122
2123 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2124
2125 /*
2126 * No need to use i_size_read() here, the i_size
2127 * cannot change under us because we hold i_mutex.
2128 *
2129 * But it's important to update i_size while still holding page lock:
2130 * page writeout could otherwise come in and zero beyond i_size.
2131 */
2132 if (pos+copied > inode->i_size) {
2133 i_size_write(inode, pos+copied);
2134 i_size_changed = 1;
2135 }
2136
2137 unlock_page(page);
2138 put_page(page);
2139
2140 if (old_size < pos)
2141 pagecache_isize_extended(inode, old_size, pos);
2142 /*
2143 * Don't mark the inode dirty under page lock. First, it unnecessarily
2144 * makes the holding time of page lock longer. Second, it forces lock
2145 * ordering of page lock and transaction start for journaling
2146 * filesystems.
2147 */
2148 if (i_size_changed)
2149 mark_inode_dirty(inode);
2150
2151 return copied;
2152}
2153EXPORT_SYMBOL(generic_write_end);
2154
2155/*
2156 * block_is_partially_uptodate checks whether buffers within a page are
2157 * uptodate or not.
2158 *
2159 * Returns true if all buffers which correspond to a file portion
2160 * we want to read are uptodate.
2161 */
2162int block_is_partially_uptodate(struct page *page, unsigned long from,
2163 unsigned long count)
2164{
2165 unsigned block_start, block_end, blocksize;
2166 unsigned to;
2167 struct buffer_head *bh, *head;
2168 int ret = 1;
2169
2170 if (!page_has_buffers(page))
2171 return 0;
2172
2173 head = page_buffers(page);
2174 blocksize = head->b_size;
2175 to = min_t(unsigned, PAGE_SIZE - from, count);
2176 to = from + to;
2177 if (from < blocksize && to > PAGE_SIZE - blocksize)
2178 return 0;
2179
2180 bh = head;
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 ret = 0;
2187 break;
2188 }
2189 if (block_end >= to)
2190 break;
2191 }
2192 block_start = block_end;
2193 bh = bh->b_this_page;
2194 } while (bh != head);
2195
2196 return ret;
2197}
2198EXPORT_SYMBOL(block_is_partially_uptodate);
2199
2200/*
2201 * Generic "read page" function for block devices that have the normal
2202 * get_block functionality. This is most of the block device filesystems.
2203 * Reads the page asynchronously --- the unlock_buffer() and
2204 * set/clear_buffer_uptodate() functions propagate buffer state into the
2205 * page struct once IO has completed.
2206 */
2207int block_read_full_page(struct page *page, get_block_t *get_block)
2208{
2209 struct inode *inode = page->mapping->host;
2210 sector_t iblock, lblock;
2211 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2212 unsigned int blocksize, bbits;
2213 int nr, i;
2214 int fully_mapped = 1;
2215
2216 head = create_page_buffers(page, inode, 0);
2217 blocksize = head->b_size;
2218 bbits = block_size_bits(blocksize);
2219
2220 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2221 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2222 bh = head;
2223 nr = 0;
2224 i = 0;
2225
2226 do {
2227 if (buffer_uptodate(bh))
2228 continue;
2229
2230 if (!buffer_mapped(bh)) {
2231 int err = 0;
2232
2233 fully_mapped = 0;
2234 if (iblock < lblock) {
2235 WARN_ON(bh->b_size != blocksize);
2236 err = get_block(inode, iblock, bh, 0);
2237 if (err)
2238 SetPageError(page);
2239 }
2240 if (!buffer_mapped(bh)) {
2241 zero_user(page, i * blocksize, blocksize);
2242 if (!err)
2243 set_buffer_uptodate(bh);
2244 continue;
2245 }
2246 /*
2247 * get_block() might have updated the buffer
2248 * synchronously
2249 */
2250 if (buffer_uptodate(bh))
2251 continue;
2252 }
2253 arr[nr++] = bh;
2254 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2255
2256 if (fully_mapped)
2257 SetPageMappedToDisk(page);
2258
2259 if (!nr) {
2260 /*
2261 * All buffers are uptodate - we can set the page uptodate
2262 * as well. But not if get_block() returned an error.
2263 */
2264 if (!PageError(page))
2265 SetPageUptodate(page);
2266 unlock_page(page);
2267 return 0;
2268 }
2269
2270 /* Stage two: lock the buffers */
2271 for (i = 0; i < nr; i++) {
2272 bh = arr[i];
2273 lock_buffer(bh);
2274 mark_buffer_async_read(bh);
2275 }
2276
2277 /*
2278 * Stage 3: start the IO. Check for uptodateness
2279 * inside the buffer lock in case another process reading
2280 * the underlying blockdev brought it uptodate (the sct fix).
2281 */
2282 for (i = 0; i < nr; i++) {
2283 bh = arr[i];
2284 if (buffer_uptodate(bh))
2285 end_buffer_async_read(bh, 1);
2286 else
2287 submit_bh(REQ_OP_READ, 0, bh);
2288 }
2289 return 0;
2290}
2291EXPORT_SYMBOL(block_read_full_page);
2292
2293/* utility function for filesystems that need to do work on expanding
2294 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2295 * deal with the hole.
2296 */
2297int generic_cont_expand_simple(struct inode *inode, loff_t size)
2298{
2299 struct address_space *mapping = inode->i_mapping;
2300 struct page *page;
2301 void *fsdata;
2302 int err;
2303
2304 err = inode_newsize_ok(inode, size);
2305 if (err)
2306 goto out;
2307
2308 err = pagecache_write_begin(NULL, mapping, size, 0,
2309 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2310 if (err)
2311 goto out;
2312
2313 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2314 BUG_ON(err > 0);
2315
2316out:
2317 return err;
2318}
2319EXPORT_SYMBOL(generic_cont_expand_simple);
2320
2321static int cont_expand_zero(struct file *file, struct address_space *mapping,
2322 loff_t pos, loff_t *bytes)
2323{
2324 struct inode *inode = mapping->host;
2325 unsigned int blocksize = i_blocksize(inode);
2326 struct page *page;
2327 void *fsdata;
2328 pgoff_t index, curidx;
2329 loff_t curpos;
2330 unsigned zerofrom, offset, len;
2331 int err = 0;
2332
2333 index = pos >> PAGE_SHIFT;
2334 offset = pos & ~PAGE_MASK;
2335
2336 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2337 zerofrom = curpos & ~PAGE_MASK;
2338 if (zerofrom & (blocksize-1)) {
2339 *bytes |= (blocksize-1);
2340 (*bytes)++;
2341 }
2342 len = PAGE_SIZE - zerofrom;
2343
2344 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2345 &page, &fsdata);
2346 if (err)
2347 goto out;
2348 zero_user(page, zerofrom, len);
2349 err = pagecache_write_end(file, mapping, curpos, len, len,
2350 page, fsdata);
2351 if (err < 0)
2352 goto out;
2353 BUG_ON(err != len);
2354 err = 0;
2355
2356 balance_dirty_pages_ratelimited(mapping);
2357
2358 if (unlikely(fatal_signal_pending(current))) {
2359 err = -EINTR;
2360 goto out;
2361 }
2362 }
2363
2364 /* page covers the boundary, find the boundary offset */
2365 if (index == curidx) {
2366 zerofrom = curpos & ~PAGE_MASK;
2367 /* if we will expand the thing last block will be filled */
2368 if (offset <= zerofrom) {
2369 goto out;
2370 }
2371 if (zerofrom & (blocksize-1)) {
2372 *bytes |= (blocksize-1);
2373 (*bytes)++;
2374 }
2375 len = offset - zerofrom;
2376
2377 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2378 &page, &fsdata);
2379 if (err)
2380 goto out;
2381 zero_user(page, zerofrom, len);
2382 err = pagecache_write_end(file, mapping, curpos, len, len,
2383 page, fsdata);
2384 if (err < 0)
2385 goto out;
2386 BUG_ON(err != len);
2387 err = 0;
2388 }
2389out:
2390 return err;
2391}
2392
2393/*
2394 * For moronic filesystems that do not allow holes in file.
2395 * We may have to extend the file.
2396 */
2397int cont_write_begin(struct file *file, struct address_space *mapping,
2398 loff_t pos, unsigned len, unsigned flags,
2399 struct page **pagep, void **fsdata,
2400 get_block_t *get_block, loff_t *bytes)
2401{
2402 struct inode *inode = mapping->host;
2403 unsigned int blocksize = i_blocksize(inode);
2404 unsigned int zerofrom;
2405 int err;
2406
2407 err = cont_expand_zero(file, mapping, pos, bytes);
2408 if (err)
2409 return err;
2410
2411 zerofrom = *bytes & ~PAGE_MASK;
2412 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2413 *bytes |= (blocksize-1);
2414 (*bytes)++;
2415 }
2416
2417 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2418}
2419EXPORT_SYMBOL(cont_write_begin);
2420
2421int block_commit_write(struct page *page, unsigned from, unsigned to)
2422{
2423 struct inode *inode = page->mapping->host;
2424 __block_commit_write(inode,page,from,to);
2425 return 0;
2426}
2427EXPORT_SYMBOL(block_commit_write);
2428
2429/*
2430 * block_page_mkwrite() is not allowed to change the file size as it gets
2431 * called from a page fault handler when a page is first dirtied. Hence we must
2432 * be careful to check for EOF conditions here. We set the page up correctly
2433 * for a written page which means we get ENOSPC checking when writing into
2434 * holes and correct delalloc and unwritten extent mapping on filesystems that
2435 * support these features.
2436 *
2437 * We are not allowed to take the i_mutex here so we have to play games to
2438 * protect against truncate races as the page could now be beyond EOF. Because
2439 * truncate writes the inode size before removing pages, once we have the
2440 * page lock we can determine safely if the page is beyond EOF. If it is not
2441 * beyond EOF, then the page is guaranteed safe against truncation until we
2442 * unlock the page.
2443 *
2444 * Direct callers of this function should protect against filesystem freezing
2445 * using sb_start_pagefault() - sb_end_pagefault() functions.
2446 */
2447int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2448 get_block_t get_block)
2449{
2450 struct page *page = vmf->page;
2451 struct inode *inode = file_inode(vma->vm_file);
2452 unsigned long end;
2453 loff_t size;
2454 int ret;
2455
2456 lock_page(page);
2457 size = i_size_read(inode);
2458 if ((page->mapping != inode->i_mapping) ||
2459 (page_offset(page) > size)) {
2460 /* We overload EFAULT to mean page got truncated */
2461 ret = -EFAULT;
2462 goto out_unlock;
2463 }
2464
2465 /* page is wholly or partially inside EOF */
2466 if (((page->index + 1) << PAGE_SHIFT) > size)
2467 end = size & ~PAGE_MASK;
2468 else
2469 end = PAGE_SIZE;
2470
2471 ret = __block_write_begin(page, 0, end, get_block);
2472 if (!ret)
2473 ret = block_commit_write(page, 0, end);
2474
2475 if (unlikely(ret < 0))
2476 goto out_unlock;
2477 set_page_dirty(page);
2478 wait_for_stable_page(page);
2479 return 0;
2480out_unlock:
2481 unlock_page(page);
2482 return ret;
2483}
2484EXPORT_SYMBOL(block_page_mkwrite);
2485
2486/*
2487 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2488 * immediately, while under the page lock. So it needs a special end_io
2489 * handler which does not touch the bh after unlocking it.
2490 */
2491static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2492{
2493 __end_buffer_read_notouch(bh, uptodate);
2494}
2495
2496/*
2497 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2498 * the page (converting it to circular linked list and taking care of page
2499 * dirty races).
2500 */
2501static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2502{
2503 struct buffer_head *bh;
2504
2505 BUG_ON(!PageLocked(page));
2506
2507 spin_lock(&page->mapping->private_lock);
2508 bh = head;
2509 do {
2510 if (PageDirty(page))
2511 set_buffer_dirty(bh);
2512 if (!bh->b_this_page)
2513 bh->b_this_page = head;
2514 bh = bh->b_this_page;
2515 } while (bh != head);
2516 attach_page_buffers(page, head);
2517 spin_unlock(&page->mapping->private_lock);
2518}
2519
2520/*
2521 * On entry, the page is fully not uptodate.
2522 * On exit the page is fully uptodate in the areas outside (from,to)
2523 * The filesystem needs to handle block truncation upon failure.
2524 */
2525int nobh_write_begin(struct address_space *mapping,
2526 loff_t pos, unsigned len, unsigned flags,
2527 struct page **pagep, void **fsdata,
2528 get_block_t *get_block)
2529{
2530 struct inode *inode = mapping->host;
2531 const unsigned blkbits = inode->i_blkbits;
2532 const unsigned blocksize = 1 << blkbits;
2533 struct buffer_head *head, *bh;
2534 struct page *page;
2535 pgoff_t index;
2536 unsigned from, to;
2537 unsigned block_in_page;
2538 unsigned block_start, block_end;
2539 sector_t block_in_file;
2540 int nr_reads = 0;
2541 int ret = 0;
2542 int is_mapped_to_disk = 1;
2543
2544 index = pos >> PAGE_SHIFT;
2545 from = pos & (PAGE_SIZE - 1);
2546 to = from + len;
2547
2548 page = grab_cache_page_write_begin(mapping, index, flags);
2549 if (!page)
2550 return -ENOMEM;
2551 *pagep = page;
2552 *fsdata = NULL;
2553
2554 if (page_has_buffers(page)) {
2555 ret = __block_write_begin(page, pos, len, get_block);
2556 if (unlikely(ret))
2557 goto out_release;
2558 return ret;
2559 }
2560
2561 if (PageMappedToDisk(page))
2562 return 0;
2563
2564 /*
2565 * Allocate buffers so that we can keep track of state, and potentially
2566 * attach them to the page if an error occurs. In the common case of
2567 * no error, they will just be freed again without ever being attached
2568 * to the page (which is all OK, because we're under the page lock).
2569 *
2570 * Be careful: the buffer linked list is a NULL terminated one, rather
2571 * than the circular one we're used to.
2572 */
2573 head = alloc_page_buffers(page, blocksize, false);
2574 if (!head) {
2575 ret = -ENOMEM;
2576 goto out_release;
2577 }
2578
2579 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2580
2581 /*
2582 * We loop across all blocks in the page, whether or not they are
2583 * part of the affected region. This is so we can discover if the
2584 * page is fully mapped-to-disk.
2585 */
2586 for (block_start = 0, block_in_page = 0, bh = head;
2587 block_start < PAGE_SIZE;
2588 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2589 int create;
2590
2591 block_end = block_start + blocksize;
2592 bh->b_state = 0;
2593 create = 1;
2594 if (block_start >= to)
2595 create = 0;
2596 ret = get_block(inode, block_in_file + block_in_page,
2597 bh, create);
2598 if (ret)
2599 goto failed;
2600 if (!buffer_mapped(bh))
2601 is_mapped_to_disk = 0;
2602 if (buffer_new(bh))
2603 clean_bdev_bh_alias(bh);
2604 if (PageUptodate(page)) {
2605 set_buffer_uptodate(bh);
2606 continue;
2607 }
2608 if (buffer_new(bh) || !buffer_mapped(bh)) {
2609 zero_user_segments(page, block_start, from,
2610 to, block_end);
2611 continue;
2612 }
2613 if (buffer_uptodate(bh))
2614 continue; /* reiserfs does this */
2615 if (block_start < from || block_end > to) {
2616 lock_buffer(bh);
2617 bh->b_end_io = end_buffer_read_nobh;
2618 submit_bh(REQ_OP_READ, 0, bh);
2619 nr_reads++;
2620 }
2621 }
2622
2623 if (nr_reads) {
2624 /*
2625 * The page is locked, so these buffers are protected from
2626 * any VM or truncate activity. Hence we don't need to care
2627 * for the buffer_head refcounts.
2628 */
2629 for (bh = head; bh; bh = bh->b_this_page) {
2630 wait_on_buffer(bh);
2631 if (!buffer_uptodate(bh))
2632 ret = -EIO;
2633 }
2634 if (ret)
2635 goto failed;
2636 }
2637
2638 if (is_mapped_to_disk)
2639 SetPageMappedToDisk(page);
2640
2641 *fsdata = head; /* to be released by nobh_write_end */
2642
2643 return 0;
2644
2645failed:
2646 BUG_ON(!ret);
2647 /*
2648 * Error recovery is a bit difficult. We need to zero out blocks that
2649 * were newly allocated, and dirty them to ensure they get written out.
2650 * Buffers need to be attached to the page at this point, otherwise
2651 * the handling of potential IO errors during writeout would be hard
2652 * (could try doing synchronous writeout, but what if that fails too?)
2653 */
2654 attach_nobh_buffers(page, head);
2655 page_zero_new_buffers(page, from, to);
2656
2657out_release:
2658 unlock_page(page);
2659 put_page(page);
2660 *pagep = NULL;
2661
2662 return ret;
2663}
2664EXPORT_SYMBOL(nobh_write_begin);
2665
2666int nobh_write_end(struct file *file, struct address_space *mapping,
2667 loff_t pos, unsigned len, unsigned copied,
2668 struct page *page, void *fsdata)
2669{
2670 struct inode *inode = page->mapping->host;
2671 struct buffer_head *head = fsdata;
2672 struct buffer_head *bh;
2673 BUG_ON(fsdata != NULL && page_has_buffers(page));
2674
2675 if (unlikely(copied < len) && head)
2676 attach_nobh_buffers(page, head);
2677 if (page_has_buffers(page))
2678 return generic_write_end(file, mapping, pos, len,
2679 copied, page, fsdata);
2680
2681 SetPageUptodate(page);
2682 set_page_dirty(page);
2683 if (pos+copied > inode->i_size) {
2684 i_size_write(inode, pos+copied);
2685 mark_inode_dirty(inode);
2686 }
2687
2688 unlock_page(page);
2689 put_page(page);
2690
2691 while (head) {
2692 bh = head;
2693 head = head->b_this_page;
2694 free_buffer_head(bh);
2695 }
2696
2697 return copied;
2698}
2699EXPORT_SYMBOL(nobh_write_end);
2700
2701/*
2702 * nobh_writepage() - based on block_full_write_page() except
2703 * that it tries to operate without attaching bufferheads to
2704 * the page.
2705 */
2706int nobh_writepage(struct page *page, get_block_t *get_block,
2707 struct writeback_control *wbc)
2708{
2709 struct inode * const inode = page->mapping->host;
2710 loff_t i_size = i_size_read(inode);
2711 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2712 unsigned offset;
2713 int ret;
2714
2715 /* Is the page fully inside i_size? */
2716 if (page->index < end_index)
2717 goto out;
2718
2719 /* Is the page fully outside i_size? (truncate in progress) */
2720 offset = i_size & (PAGE_SIZE-1);
2721 if (page->index >= end_index+1 || !offset) {
2722 /*
2723 * The page may have dirty, unmapped buffers. For example,
2724 * they may have been added in ext3_writepage(). Make them
2725 * freeable here, so the page does not leak.
2726 */
2727#if 0
2728 /* Not really sure about this - do we need this ? */
2729 if (page->mapping->a_ops->invalidatepage)
2730 page->mapping->a_ops->invalidatepage(page, offset);
2731#endif
2732 unlock_page(page);
2733 return 0; /* don't care */
2734 }
2735
2736 /*
2737 * The page straddles i_size. It must be zeroed out on each and every
2738 * writepage invocation because it may be mmapped. "A file is mapped
2739 * in multiples of the page size. For a file that is not a multiple of
2740 * the page size, the remaining memory is zeroed when mapped, and
2741 * writes to that region are not written out to the file."
2742 */
2743 zero_user_segment(page, offset, PAGE_SIZE);
2744out:
2745 ret = mpage_writepage(page, get_block, wbc);
2746 if (ret == -EAGAIN)
2747 ret = __block_write_full_page(inode, page, get_block, wbc,
2748 end_buffer_async_write);
2749 return ret;
2750}
2751EXPORT_SYMBOL(nobh_writepage);
2752
2753int nobh_truncate_page(struct address_space *mapping,
2754 loff_t from, get_block_t *get_block)
2755{
2756 pgoff_t index = from >> PAGE_SHIFT;
2757 unsigned offset = from & (PAGE_SIZE-1);
2758 unsigned blocksize;
2759 sector_t iblock;
2760 unsigned length, pos;
2761 struct inode *inode = mapping->host;
2762 struct page *page;
2763 struct buffer_head map_bh;
2764 int err;
2765
2766 blocksize = i_blocksize(inode);
2767 length = offset & (blocksize - 1);
2768
2769 /* Block boundary? Nothing to do */
2770 if (!length)
2771 return 0;
2772
2773 length = blocksize - length;
2774 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2775
2776 page = grab_cache_page(mapping, index);
2777 err = -ENOMEM;
2778 if (!page)
2779 goto out;
2780
2781 if (page_has_buffers(page)) {
2782has_buffers:
2783 unlock_page(page);
2784 put_page(page);
2785 return block_truncate_page(mapping, from, get_block);
2786 }
2787
2788 /* Find the buffer that contains "offset" */
2789 pos = blocksize;
2790 while (offset >= pos) {
2791 iblock++;
2792 pos += blocksize;
2793 }
2794
2795 map_bh.b_size = blocksize;
2796 map_bh.b_state = 0;
2797 err = get_block(inode, iblock, &map_bh, 0);
2798 if (err)
2799 goto unlock;
2800 /* unmapped? It's a hole - nothing to do */
2801 if (!buffer_mapped(&map_bh))
2802 goto unlock;
2803
2804 /* Ok, it's mapped. Make sure it's up-to-date */
2805 if (!PageUptodate(page)) {
2806 err = mapping->a_ops->readpage(NULL, page);
2807 if (err) {
2808 put_page(page);
2809 goto out;
2810 }
2811 lock_page(page);
2812 if (!PageUptodate(page)) {
2813 err = -EIO;
2814 goto unlock;
2815 }
2816 if (page_has_buffers(page))
2817 goto has_buffers;
2818 }
2819 zero_user(page, offset, length);
2820 set_page_dirty(page);
2821 err = 0;
2822
2823unlock:
2824 unlock_page(page);
2825 put_page(page);
2826out:
2827 return err;
2828}
2829EXPORT_SYMBOL(nobh_truncate_page);
2830
2831int block_truncate_page(struct address_space *mapping,
2832 loff_t from, get_block_t *get_block)
2833{
2834 pgoff_t index = from >> PAGE_SHIFT;
2835 unsigned offset = from & (PAGE_SIZE-1);
2836 unsigned blocksize;
2837 sector_t iblock;
2838 unsigned length, pos;
2839 struct inode *inode = mapping->host;
2840 struct page *page;
2841 struct buffer_head *bh;
2842 int err;
2843
2844 blocksize = i_blocksize(inode);
2845 length = offset & (blocksize - 1);
2846
2847 /* Block boundary? Nothing to do */
2848 if (!length)
2849 return 0;
2850
2851 length = blocksize - length;
2852 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2853
2854 page = grab_cache_page(mapping, index);
2855 err = -ENOMEM;
2856 if (!page)
2857 goto out;
2858
2859 if (!page_has_buffers(page))
2860 create_empty_buffers(page, blocksize, 0);
2861
2862 /* Find the buffer that contains "offset" */
2863 bh = page_buffers(page);
2864 pos = blocksize;
2865 while (offset >= pos) {
2866 bh = bh->b_this_page;
2867 iblock++;
2868 pos += blocksize;
2869 }
2870
2871 err = 0;
2872 if (!buffer_mapped(bh)) {
2873 WARN_ON(bh->b_size != blocksize);
2874 err = get_block(inode, iblock, bh, 0);
2875 if (err)
2876 goto unlock;
2877 /* unmapped? It's a hole - nothing to do */
2878 if (!buffer_mapped(bh))
2879 goto unlock;
2880 }
2881
2882 /* Ok, it's mapped. Make sure it's up-to-date */
2883 if (PageUptodate(page))
2884 set_buffer_uptodate(bh);
2885
2886 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2887 err = -EIO;
2888 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2889 wait_on_buffer(bh);
2890 /* Uhhuh. Read error. Complain and punt. */
2891 if (!buffer_uptodate(bh))
2892 goto unlock;
2893 }
2894
2895 zero_user(page, offset, length);
2896 mark_buffer_dirty(bh);
2897 err = 0;
2898
2899unlock:
2900 unlock_page(page);
2901 put_page(page);
2902out:
2903 return err;
2904}
2905EXPORT_SYMBOL(block_truncate_page);
2906
2907/*
2908 * The generic ->writepage function for buffer-backed address_spaces
2909 */
2910int block_write_full_page(struct page *page, get_block_t *get_block,
2911 struct writeback_control *wbc)
2912{
2913 struct inode * const inode = page->mapping->host;
2914 loff_t i_size = i_size_read(inode);
2915 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2916 unsigned offset;
2917
2918 /* Is the page fully inside i_size? */
2919 if (page->index < end_index)
2920 return __block_write_full_page(inode, page, get_block, wbc,
2921 end_buffer_async_write);
2922
2923 /* Is the page fully outside i_size? (truncate in progress) */
2924 offset = i_size & (PAGE_SIZE-1);
2925 if (page->index >= end_index+1 || !offset) {
2926 /*
2927 * The page may have dirty, unmapped buffers. For example,
2928 * they may have been added in ext3_writepage(). Make them
2929 * freeable here, so the page does not leak.
2930 */
2931 do_invalidatepage(page, 0, PAGE_SIZE);
2932 unlock_page(page);
2933 return 0; /* don't care */
2934 }
2935
2936 /*
2937 * The page straddles i_size. It must be zeroed out on each and every
2938 * writepage invocation because it may be mmapped. "A file is mapped
2939 * in multiples of the page size. For a file that is not a multiple of
2940 * the page size, the remaining memory is zeroed when mapped, and
2941 * writes to that region are not written out to the file."
2942 */
2943 zero_user_segment(page, offset, PAGE_SIZE);
2944 return __block_write_full_page(inode, page, get_block, wbc,
2945 end_buffer_async_write);
2946}
2947EXPORT_SYMBOL(block_write_full_page);
2948
2949sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2950 get_block_t *get_block)
2951{
2952 struct inode *inode = mapping->host;
2953 struct buffer_head tmp = {
2954 .b_size = i_blocksize(inode),
2955 };
2956
2957 get_block(inode, block, &tmp, 0);
2958 return tmp.b_blocknr;
2959}
2960EXPORT_SYMBOL(generic_block_bmap);
2961
2962static void end_bio_bh_io_sync(struct bio *bio)
2963{
2964 struct buffer_head *bh = bio->bi_private;
2965
2966 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2967 set_bit(BH_Quiet, &bh->b_state);
2968
2969 bh->b_end_io(bh, !bio->bi_status);
2970 bio_put(bio);
2971}
2972
2973/*
2974 * This allows us to do IO even on the odd last sectors
2975 * of a device, even if the block size is some multiple
2976 * of the physical sector size.
2977 *
2978 * We'll just truncate the bio to the size of the device,
2979 * and clear the end of the buffer head manually.
2980 *
2981 * Truly out-of-range accesses will turn into actual IO
2982 * errors, this only handles the "we need to be able to
2983 * do IO at the final sector" case.
2984 */
2985void guard_bio_eod(int op, struct bio *bio)
2986{
2987 sector_t maxsector;
2988 struct bio_vec *bvec = bio_last_bvec_all(bio);
2989 unsigned truncated_bytes;
2990 struct hd_struct *part;
2991
2992 rcu_read_lock();
2993 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
2994 if (part)
2995 maxsector = part_nr_sects_read(part);
2996 else
2997 maxsector = get_capacity(bio->bi_disk);
2998 rcu_read_unlock();
2999
3000 if (!maxsector)
3001 return;
3002
3003 /*
3004 * If the *whole* IO is past the end of the device,
3005 * let it through, and the IO layer will turn it into
3006 * an EIO.
3007 */
3008 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3009 return;
3010
3011 maxsector -= bio->bi_iter.bi_sector;
3012 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3013 return;
3014
3015 /* Uhhuh. We've got a bio that straddles the device size! */
3016 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3017
3018 /* Truncate the bio.. */
3019 bio->bi_iter.bi_size -= truncated_bytes;
3020 bvec->bv_len -= truncated_bytes;
3021
3022 /* ..and clear the end of the buffer for reads */
3023 if (op == REQ_OP_READ) {
3024 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3025 truncated_bytes);
3026 }
3027}
3028
3029static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3030 enum rw_hint write_hint, struct writeback_control *wbc)
3031{
3032 struct bio *bio;
3033
3034 BUG_ON(!buffer_locked(bh));
3035 BUG_ON(!buffer_mapped(bh));
3036 BUG_ON(!bh->b_end_io);
3037 BUG_ON(buffer_delay(bh));
3038 BUG_ON(buffer_unwritten(bh));
3039
3040 /*
3041 * Only clear out a write error when rewriting
3042 */
3043 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3044 clear_buffer_write_io_error(bh);
3045
3046 /*
3047 * from here on down, it's all bio -- do the initial mapping,
3048 * submit_bio -> generic_make_request may further map this bio around
3049 */
3050 bio = bio_alloc(GFP_NOIO, 1);
3051
3052 if (wbc) {
3053 wbc_init_bio(wbc, bio);
3054 wbc_account_io(wbc, bh->b_page, bh->b_size);
3055 }
3056
3057 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3058 bio_set_dev(bio, bh->b_bdev);
3059 bio->bi_write_hint = write_hint;
3060
3061 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3062 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3063
3064 bio->bi_end_io = end_bio_bh_io_sync;
3065 bio->bi_private = bh;
3066
3067 /* Take care of bh's that straddle the end of the device */
3068 guard_bio_eod(op, bio);
3069
3070 if (buffer_meta(bh))
3071 op_flags |= REQ_META;
3072 if (buffer_prio(bh))
3073 op_flags |= REQ_PRIO;
3074 bio_set_op_attrs(bio, op, op_flags);
3075
3076 submit_bio(bio);
3077 return 0;
3078}
3079
3080int submit_bh(int op, int op_flags, struct buffer_head *bh)
3081{
3082 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3083}
3084EXPORT_SYMBOL(submit_bh);
3085
3086/**
3087 * ll_rw_block: low-level access to block devices (DEPRECATED)
3088 * @op: whether to %READ or %WRITE
3089 * @op_flags: req_flag_bits
3090 * @nr: number of &struct buffer_heads in the array
3091 * @bhs: array of pointers to &struct buffer_head
3092 *
3093 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3094 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3095 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3096 * %REQ_RAHEAD.
3097 *
3098 * This function drops any buffer that it cannot get a lock on (with the
3099 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3100 * request, and any buffer that appears to be up-to-date when doing read
3101 * request. Further it marks as clean buffers that are processed for
3102 * writing (the buffer cache won't assume that they are actually clean
3103 * until the buffer gets unlocked).
3104 *
3105 * ll_rw_block sets b_end_io to simple completion handler that marks
3106 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3107 * any waiters.
3108 *
3109 * All of the buffers must be for the same device, and must also be a
3110 * multiple of the current approved size for the device.
3111 */
3112void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3113{
3114 int i;
3115
3116 for (i = 0; i < nr; i++) {
3117 struct buffer_head *bh = bhs[i];
3118
3119 if (!trylock_buffer(bh))
3120 continue;
3121 if (op == WRITE) {
3122 if (test_clear_buffer_dirty(bh)) {
3123 bh->b_end_io = end_buffer_write_sync;
3124 get_bh(bh);
3125 submit_bh(op, op_flags, bh);
3126 continue;
3127 }
3128 } else {
3129 if (!buffer_uptodate(bh)) {
3130 bh->b_end_io = end_buffer_read_sync;
3131 get_bh(bh);
3132 submit_bh(op, op_flags, bh);
3133 continue;
3134 }
3135 }
3136 unlock_buffer(bh);
3137 }
3138}
3139EXPORT_SYMBOL(ll_rw_block);
3140
3141void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3142{
3143 lock_buffer(bh);
3144 if (!test_clear_buffer_dirty(bh)) {
3145 unlock_buffer(bh);
3146 return;
3147 }
3148 bh->b_end_io = end_buffer_write_sync;
3149 get_bh(bh);
3150 submit_bh(REQ_OP_WRITE, op_flags, bh);
3151}
3152EXPORT_SYMBOL(write_dirty_buffer);
3153
3154/*
3155 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3156 * and then start new I/O and then wait upon it. The caller must have a ref on
3157 * the buffer_head.
3158 */
3159int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3160{
3161 int ret = 0;
3162
3163 WARN_ON(atomic_read(&bh->b_count) < 1);
3164 lock_buffer(bh);
3165 if (test_clear_buffer_dirty(bh)) {
3166 get_bh(bh);
3167 bh->b_end_io = end_buffer_write_sync;
3168 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3169 wait_on_buffer(bh);
3170 if (!ret && !buffer_uptodate(bh))
3171 ret = -EIO;
3172 } else {
3173 unlock_buffer(bh);
3174 }
3175 return ret;
3176}
3177EXPORT_SYMBOL(__sync_dirty_buffer);
3178
3179int sync_dirty_buffer(struct buffer_head *bh)
3180{
3181 return __sync_dirty_buffer(bh, REQ_SYNC);
3182}
3183EXPORT_SYMBOL(sync_dirty_buffer);
3184
3185/*
3186 * try_to_free_buffers() checks if all the buffers on this particular page
3187 * are unused, and releases them if so.
3188 *
3189 * Exclusion against try_to_free_buffers may be obtained by either
3190 * locking the page or by holding its mapping's private_lock.
3191 *
3192 * If the page is dirty but all the buffers are clean then we need to
3193 * be sure to mark the page clean as well. This is because the page
3194 * may be against a block device, and a later reattachment of buffers
3195 * to a dirty page will set *all* buffers dirty. Which would corrupt
3196 * filesystem data on the same device.
3197 *
3198 * The same applies to regular filesystem pages: if all the buffers are
3199 * clean then we set the page clean and proceed. To do that, we require
3200 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3201 * private_lock.
3202 *
3203 * try_to_free_buffers() is non-blocking.
3204 */
3205static inline int buffer_busy(struct buffer_head *bh)
3206{
3207 return atomic_read(&bh->b_count) |
3208 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3209}
3210
3211static int
3212drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3213{
3214 struct buffer_head *head = page_buffers(page);
3215 struct buffer_head *bh;
3216
3217 bh = head;
3218 do {
3219 if (buffer_busy(bh))
3220 goto failed;
3221 bh = bh->b_this_page;
3222 } while (bh != head);
3223
3224 do {
3225 struct buffer_head *next = bh->b_this_page;
3226
3227 if (bh->b_assoc_map)
3228 __remove_assoc_queue(bh);
3229 bh = next;
3230 } while (bh != head);
3231 *buffers_to_free = head;
3232 __clear_page_buffers(page);
3233 return 1;
3234failed:
3235 return 0;
3236}
3237
3238int try_to_free_buffers(struct page *page)
3239{
3240 struct address_space * const mapping = page->mapping;
3241 struct buffer_head *buffers_to_free = NULL;
3242 int ret = 0;
3243
3244 BUG_ON(!PageLocked(page));
3245 if (PageWriteback(page))
3246 return 0;
3247
3248 if (mapping == NULL) { /* can this still happen? */
3249 ret = drop_buffers(page, &buffers_to_free);
3250 goto out;
3251 }
3252
3253 spin_lock(&mapping->private_lock);
3254 ret = drop_buffers(page, &buffers_to_free);
3255
3256 /*
3257 * If the filesystem writes its buffers by hand (eg ext3)
3258 * then we can have clean buffers against a dirty page. We
3259 * clean the page here; otherwise the VM will never notice
3260 * that the filesystem did any IO at all.
3261 *
3262 * Also, during truncate, discard_buffer will have marked all
3263 * the page's buffers clean. We discover that here and clean
3264 * the page also.
3265 *
3266 * private_lock must be held over this entire operation in order
3267 * to synchronise against __set_page_dirty_buffers and prevent the
3268 * dirty bit from being lost.
3269 */
3270 if (ret)
3271 cancel_dirty_page(page);
3272 spin_unlock(&mapping->private_lock);
3273out:
3274 if (buffers_to_free) {
3275 struct buffer_head *bh = buffers_to_free;
3276
3277 do {
3278 struct buffer_head *next = bh->b_this_page;
3279 free_buffer_head(bh);
3280 bh = next;
3281 } while (bh != buffers_to_free);
3282 }
3283 return ret;
3284}
3285EXPORT_SYMBOL(try_to_free_buffers);
3286
3287/*
3288 * There are no bdflush tunables left. But distributions are
3289 * still running obsolete flush daemons, so we terminate them here.
3290 *
3291 * Use of bdflush() is deprecated and will be removed in a future kernel.
3292 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3293 */
3294SYSCALL_DEFINE2(bdflush, int, func, long, data)
3295{
3296 static int msg_count;
3297
3298 if (!capable(CAP_SYS_ADMIN))
3299 return -EPERM;
3300
3301 if (msg_count < 5) {
3302 msg_count++;
3303 printk(KERN_INFO
3304 "warning: process `%s' used the obsolete bdflush"
3305 " system call\n", current->comm);
3306 printk(KERN_INFO "Fix your initscripts?\n");
3307 }
3308
3309 if (func == 1)
3310 do_exit(0);
3311 return 0;
3312}
3313
3314/*
3315 * Buffer-head allocation
3316 */
3317static struct kmem_cache *bh_cachep __read_mostly;
3318
3319/*
3320 * Once the number of bh's in the machine exceeds this level, we start
3321 * stripping them in writeback.
3322 */
3323static unsigned long max_buffer_heads;
3324
3325int buffer_heads_over_limit;
3326
3327struct bh_accounting {
3328 int nr; /* Number of live bh's */
3329 int ratelimit; /* Limit cacheline bouncing */
3330};
3331
3332static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3333
3334static void recalc_bh_state(void)
3335{
3336 int i;
3337 int tot = 0;
3338
3339 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3340 return;
3341 __this_cpu_write(bh_accounting.ratelimit, 0);
3342 for_each_online_cpu(i)
3343 tot += per_cpu(bh_accounting, i).nr;
3344 buffer_heads_over_limit = (tot > max_buffer_heads);
3345}
3346
3347struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3348{
3349 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3350 if (ret) {
3351 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3352 preempt_disable();
3353 __this_cpu_inc(bh_accounting.nr);
3354 recalc_bh_state();
3355 preempt_enable();
3356 }
3357 return ret;
3358}
3359EXPORT_SYMBOL(alloc_buffer_head);
3360
3361void free_buffer_head(struct buffer_head *bh)
3362{
3363 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3364 kmem_cache_free(bh_cachep, bh);
3365 preempt_disable();
3366 __this_cpu_dec(bh_accounting.nr);
3367 recalc_bh_state();
3368 preempt_enable();
3369}
3370EXPORT_SYMBOL(free_buffer_head);
3371
3372static int buffer_exit_cpu_dead(unsigned int cpu)
3373{
3374 int i;
3375 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3376
3377 for (i = 0; i < BH_LRU_SIZE; i++) {
3378 brelse(b->bhs[i]);
3379 b->bhs[i] = NULL;
3380 }
3381 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3382 per_cpu(bh_accounting, cpu).nr = 0;
3383 return 0;
3384}
3385
3386/**
3387 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3388 * @bh: struct buffer_head
3389 *
3390 * Return true if the buffer is up-to-date and false,
3391 * with the buffer locked, if not.
3392 */
3393int bh_uptodate_or_lock(struct buffer_head *bh)
3394{
3395 if (!buffer_uptodate(bh)) {
3396 lock_buffer(bh);
3397 if (!buffer_uptodate(bh))
3398 return 0;
3399 unlock_buffer(bh);
3400 }
3401 return 1;
3402}
3403EXPORT_SYMBOL(bh_uptodate_or_lock);
3404
3405/**
3406 * bh_submit_read - Submit a locked buffer for reading
3407 * @bh: struct buffer_head
3408 *
3409 * Returns zero on success and -EIO on error.
3410 */
3411int bh_submit_read(struct buffer_head *bh)
3412{
3413 BUG_ON(!buffer_locked(bh));
3414
3415 if (buffer_uptodate(bh)) {
3416 unlock_buffer(bh);
3417 return 0;
3418 }
3419
3420 get_bh(bh);
3421 bh->b_end_io = end_buffer_read_sync;
3422 submit_bh(REQ_OP_READ, 0, bh);
3423 wait_on_buffer(bh);
3424 if (buffer_uptodate(bh))
3425 return 0;
3426 return -EIO;
3427}
3428EXPORT_SYMBOL(bh_submit_read);
3429
3430/*
3431 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
3432 *
3433 * Returns the offset within the file on success, and -ENOENT otherwise.
3434 */
3435static loff_t
3436page_seek_hole_data(struct page *page, loff_t lastoff, int whence)
3437{
3438 loff_t offset = page_offset(page);
3439 struct buffer_head *bh, *head;
3440 bool seek_data = whence == SEEK_DATA;
3441
3442 if (lastoff < offset)
3443 lastoff = offset;
3444
3445 bh = head = page_buffers(page);
3446 do {
3447 offset += bh->b_size;
3448 if (lastoff >= offset)
3449 continue;
3450
3451 /*
3452 * Unwritten extents that have data in the page cache covering
3453 * them can be identified by the BH_Unwritten state flag.
3454 * Pages with multiple buffers might have a mix of holes, data
3455 * and unwritten extents - any buffer with valid data in it
3456 * should have BH_Uptodate flag set on it.
3457 */
3458
3459 if ((buffer_unwritten(bh) || buffer_uptodate(bh)) == seek_data)
3460 return lastoff;
3461
3462 lastoff = offset;
3463 } while ((bh = bh->b_this_page) != head);
3464 return -ENOENT;
3465}
3466
3467/*
3468 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3469 *
3470 * Within unwritten extents, the page cache determines which parts are holes
3471 * and which are data: unwritten and uptodate buffer heads count as data;
3472 * everything else counts as a hole.
3473 *
3474 * Returns the resulting offset on successs, and -ENOENT otherwise.
3475 */
3476loff_t
3477page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
3478 int whence)
3479{
3480 pgoff_t index = offset >> PAGE_SHIFT;
3481 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
3482 loff_t lastoff = offset;
3483 struct pagevec pvec;
3484
3485 if (length <= 0)
3486 return -ENOENT;
3487
3488 pagevec_init(&pvec);
3489
3490 do {
3491 unsigned nr_pages, i;
3492
3493 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
3494 end - 1);
3495 if (nr_pages == 0)
3496 break;
3497
3498 for (i = 0; i < nr_pages; i++) {
3499 struct page *page = pvec.pages[i];
3500
3501 /*
3502 * At this point, the page may be truncated or
3503 * invalidated (changing page->mapping to NULL), or
3504 * even swizzled back from swapper_space to tmpfs file
3505 * mapping. However, page->index will not change
3506 * because we have a reference on the page.
3507 *
3508 * If current page offset is beyond where we've ended,
3509 * we've found a hole.
3510 */
3511 if (whence == SEEK_HOLE &&
3512 lastoff < page_offset(page))
3513 goto check_range;
3514
3515 lock_page(page);
3516 if (likely(page->mapping == inode->i_mapping) &&
3517 page_has_buffers(page)) {
3518 lastoff = page_seek_hole_data(page, lastoff, whence);
3519 if (lastoff >= 0) {
3520 unlock_page(page);
3521 goto check_range;
3522 }
3523 }
3524 unlock_page(page);
3525 lastoff = page_offset(page) + PAGE_SIZE;
3526 }
3527 pagevec_release(&pvec);
3528 } while (index < end);
3529
3530 /* When no page at lastoff and we are not done, we found a hole. */
3531 if (whence != SEEK_HOLE)
3532 goto not_found;
3533
3534check_range:
3535 if (lastoff < offset + length)
3536 goto out;
3537not_found:
3538 lastoff = -ENOENT;
3539out:
3540 pagevec_release(&pvec);
3541 return lastoff;
3542}
3543
3544void __init buffer_init(void)
3545{
3546 unsigned long nrpages;
3547 int ret;
3548
3549 bh_cachep = kmem_cache_create("buffer_head",
3550 sizeof(struct buffer_head), 0,
3551 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3552 SLAB_MEM_SPREAD),
3553 NULL);
3554
3555 /*
3556 * Limit the bh occupancy to 10% of ZONE_NORMAL
3557 */
3558 nrpages = (nr_free_buffer_pages() * 10) / 100;
3559 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3560 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3561 NULL, buffer_exit_cpu_dead);
3562 WARN_ON(ret < 0);
3563}