1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2002, Linus Torvalds.
7 * Contains functions related to preparing and submitting BIOs which contain
8 * multiple pagecache pages.
10 * 15May2002 Andrew Morton
12 * 27Jun2002 axboe@suse.de
13 * use bio_add_page() to build bio's just the right size
16 #include <linux/kernel.h>
17 #include <linux/export.h>
19 #include <linux/kdev_t.h>
20 #include <linux/gfp.h>
21 #include <linux/bio.h>
23 #include <linux/buffer_head.h>
24 #include <linux/blkdev.h>
25 #include <linux/highmem.h>
26 #include <linux/prefetch.h>
27 #include <linux/mpage.h>
28 #include <linux/mm_inline.h>
29 #include <linux/writeback.h>
30 #include <linux/backing-dev.h>
31 #include <linux/pagevec.h>
35 * I/O completion handler for multipage BIOs.
37 * The mpage code never puts partial pages into a BIO (except for end-of-file).
38 * If a page does not map to a contiguous run of blocks then it simply falls
39 * back to block_read_full_folio().
41 * Why is this? If a page's completion depends on a number of different BIOs
42 * which can complete in any order (or at the same time) then determining the
43 * status of that page is hard. See end_buffer_async_read() for the details.
44 * There is no point in duplicating all that complexity.
46 static void mpage_read_end_io(struct bio
*bio
)
49 int err
= blk_status_to_errno(bio
->bi_status
);
51 bio_for_each_folio_all(fi
, bio
) {
53 folio_set_error(fi
.folio
);
55 folio_mark_uptodate(fi
.folio
);
56 folio_unlock(fi
.folio
);
62 static void mpage_write_end_io(struct bio
*bio
)
65 int err
= blk_status_to_errno(bio
->bi_status
);
67 bio_for_each_folio_all(fi
, bio
) {
69 folio_set_error(fi
.folio
);
70 mapping_set_error(fi
.folio
->mapping
, err
);
72 folio_end_writeback(fi
.folio
);
78 static struct bio
*mpage_bio_submit_read(struct bio
*bio
)
80 bio
->bi_end_io
= mpage_read_end_io
;
86 static struct bio
*mpage_bio_submit_write(struct bio
*bio
)
88 bio
->bi_end_io
= mpage_write_end_io
;
95 * support function for mpage_readahead. The fs supplied get_block might
96 * return an up to date buffer. This is used to map that buffer into
97 * the page, which allows read_folio to avoid triggering a duplicate call
100 * The idea is to avoid adding buffers to pages that don't already have
101 * them. So when the buffer is up to date and the page size == block size,
102 * this marks the page up to date instead of adding new buffers.
104 static void map_buffer_to_folio(struct folio
*folio
, struct buffer_head
*bh
,
107 struct inode
*inode
= folio
->mapping
->host
;
108 struct buffer_head
*page_bh
, *head
;
111 head
= folio_buffers(folio
);
114 * don't make any buffers if there is only one buffer on
115 * the folio and the folio just needs to be set up to date
117 if (inode
->i_blkbits
== PAGE_SHIFT
&&
118 buffer_uptodate(bh
)) {
119 folio_mark_uptodate(folio
);
122 head
= create_empty_buffers(folio
, i_blocksize(inode
), 0);
127 if (block
== page_block
) {
128 page_bh
->b_state
= bh
->b_state
;
129 page_bh
->b_bdev
= bh
->b_bdev
;
130 page_bh
->b_blocknr
= bh
->b_blocknr
;
133 page_bh
= page_bh
->b_this_page
;
135 } while (page_bh
!= head
);
138 struct mpage_readpage_args
{
141 unsigned int nr_pages
;
143 sector_t last_block_in_bio
;
144 struct buffer_head map_bh
;
145 unsigned long first_logical_block
;
146 get_block_t
*get_block
;
150 * This is the worker routine which does all the work of mapping the disk
151 * blocks and constructs largest possible bios, submits them for IO if the
152 * blocks are not contiguous on the disk.
154 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
155 * represent the validity of its disk mapping and to decide when to do the next
158 static struct bio
*do_mpage_readpage(struct mpage_readpage_args
*args
)
160 struct folio
*folio
= args
->folio
;
161 struct inode
*inode
= folio
->mapping
->host
;
162 const unsigned blkbits
= inode
->i_blkbits
;
163 const unsigned blocks_per_page
= PAGE_SIZE
>> blkbits
;
164 const unsigned blocksize
= 1 << blkbits
;
165 struct buffer_head
*map_bh
= &args
->map_bh
;
166 sector_t block_in_file
;
168 sector_t last_block_in_file
;
169 sector_t blocks
[MAX_BUF_PER_PAGE
];
171 unsigned first_hole
= blocks_per_page
;
172 struct block_device
*bdev
= NULL
;
174 int fully_mapped
= 1;
175 blk_opf_t opf
= REQ_OP_READ
;
177 unsigned relative_block
;
178 gfp_t gfp
= mapping_gfp_constraint(folio
->mapping
, GFP_KERNEL
);
180 /* MAX_BUF_PER_PAGE, for example */
181 VM_BUG_ON_FOLIO(folio_test_large(folio
), folio
);
183 if (args
->is_readahead
) {
185 gfp
|= __GFP_NORETRY
| __GFP_NOWARN
;
188 if (folio_buffers(folio
))
191 block_in_file
= (sector_t
)folio
->index
<< (PAGE_SHIFT
- blkbits
);
192 last_block
= block_in_file
+ args
->nr_pages
* blocks_per_page
;
193 last_block_in_file
= (i_size_read(inode
) + blocksize
- 1) >> blkbits
;
194 if (last_block
> last_block_in_file
)
195 last_block
= last_block_in_file
;
199 * Map blocks using the result from the previous get_blocks call first.
201 nblocks
= map_bh
->b_size
>> blkbits
;
202 if (buffer_mapped(map_bh
) &&
203 block_in_file
> args
->first_logical_block
&&
204 block_in_file
< (args
->first_logical_block
+ nblocks
)) {
205 unsigned map_offset
= block_in_file
- args
->first_logical_block
;
206 unsigned last
= nblocks
- map_offset
;
208 for (relative_block
= 0; ; relative_block
++) {
209 if (relative_block
== last
) {
210 clear_buffer_mapped(map_bh
);
213 if (page_block
== blocks_per_page
)
215 blocks
[page_block
] = map_bh
->b_blocknr
+ map_offset
+
220 bdev
= map_bh
->b_bdev
;
224 * Then do more get_blocks calls until we are done with this folio.
226 map_bh
->b_folio
= folio
;
227 while (page_block
< blocks_per_page
) {
231 if (block_in_file
< last_block
) {
232 map_bh
->b_size
= (last_block
-block_in_file
) << blkbits
;
233 if (args
->get_block(inode
, block_in_file
, map_bh
, 0))
235 args
->first_logical_block
= block_in_file
;
238 if (!buffer_mapped(map_bh
)) {
240 if (first_hole
== blocks_per_page
)
241 first_hole
= page_block
;
247 /* some filesystems will copy data into the page during
248 * the get_block call, in which case we don't want to
249 * read it again. map_buffer_to_folio copies the data
250 * we just collected from get_block into the folio's buffers
251 * so read_folio doesn't have to repeat the get_block call
253 if (buffer_uptodate(map_bh
)) {
254 map_buffer_to_folio(folio
, map_bh
, page_block
);
258 if (first_hole
!= blocks_per_page
)
259 goto confused
; /* hole -> non-hole */
261 /* Contiguous blocks? */
262 if (page_block
&& blocks
[page_block
-1] != map_bh
->b_blocknr
-1)
264 nblocks
= map_bh
->b_size
>> blkbits
;
265 for (relative_block
= 0; ; relative_block
++) {
266 if (relative_block
== nblocks
) {
267 clear_buffer_mapped(map_bh
);
269 } else if (page_block
== blocks_per_page
)
271 blocks
[page_block
] = map_bh
->b_blocknr
+relative_block
;
275 bdev
= map_bh
->b_bdev
;
278 if (first_hole
!= blocks_per_page
) {
279 folio_zero_segment(folio
, first_hole
<< blkbits
, PAGE_SIZE
);
280 if (first_hole
== 0) {
281 folio_mark_uptodate(folio
);
285 } else if (fully_mapped
) {
286 folio_set_mappedtodisk(folio
);
290 * This folio will go to BIO. Do we need to send this BIO off first?
292 if (args
->bio
&& (args
->last_block_in_bio
!= blocks
[0] - 1))
293 args
->bio
= mpage_bio_submit_read(args
->bio
);
296 if (args
->bio
== NULL
) {
297 args
->bio
= bio_alloc(bdev
, bio_max_segs(args
->nr_pages
), opf
,
299 if (args
->bio
== NULL
)
301 args
->bio
->bi_iter
.bi_sector
= blocks
[0] << (blkbits
- 9);
304 length
= first_hole
<< blkbits
;
305 if (!bio_add_folio(args
->bio
, folio
, length
, 0)) {
306 args
->bio
= mpage_bio_submit_read(args
->bio
);
310 relative_block
= block_in_file
- args
->first_logical_block
;
311 nblocks
= map_bh
->b_size
>> blkbits
;
312 if ((buffer_boundary(map_bh
) && relative_block
== nblocks
) ||
313 (first_hole
!= blocks_per_page
))
314 args
->bio
= mpage_bio_submit_read(args
->bio
);
316 args
->last_block_in_bio
= blocks
[blocks_per_page
- 1];
322 args
->bio
= mpage_bio_submit_read(args
->bio
);
323 if (!folio_test_uptodate(folio
))
324 block_read_full_folio(folio
, args
->get_block
);
331 * mpage_readahead - start reads against pages
332 * @rac: Describes which pages to read.
333 * @get_block: The filesystem's block mapper function.
335 * This function walks the pages and the blocks within each page, building and
336 * emitting large BIOs.
338 * If anything unusual happens, such as:
340 * - encountering a page which has buffers
341 * - encountering a page which has a non-hole after a hole
342 * - encountering a page with non-contiguous blocks
344 * then this code just gives up and calls the buffer_head-based read function.
345 * It does handle a page which has holes at the end - that is a common case:
346 * the end-of-file on blocksize < PAGE_SIZE setups.
348 * BH_Boundary explanation:
350 * There is a problem. The mpage read code assembles several pages, gets all
351 * their disk mappings, and then submits them all. That's fine, but obtaining
352 * the disk mappings may require I/O. Reads of indirect blocks, for example.
354 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
355 * submitted in the following order:
357 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
359 * because the indirect block has to be read to get the mappings of blocks
360 * 13,14,15,16. Obviously, this impacts performance.
362 * So what we do it to allow the filesystem's get_block() function to set
363 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
364 * after this one will require I/O against a block which is probably close to
365 * this one. So you should push what I/O you have currently accumulated.
367 * This all causes the disk requests to be issued in the correct order.
369 void mpage_readahead(struct readahead_control
*rac
, get_block_t get_block
)
372 struct mpage_readpage_args args
= {
373 .get_block
= get_block
,
374 .is_readahead
= true,
377 while ((folio
= readahead_folio(rac
))) {
378 prefetchw(&folio
->flags
);
380 args
.nr_pages
= readahead_count(rac
);
381 args
.bio
= do_mpage_readpage(&args
);
384 mpage_bio_submit_read(args
.bio
);
386 EXPORT_SYMBOL(mpage_readahead
);
389 * This isn't called much at all
391 int mpage_read_folio(struct folio
*folio
, get_block_t get_block
)
393 struct mpage_readpage_args args
= {
396 .get_block
= get_block
,
399 args
.bio
= do_mpage_readpage(&args
);
401 mpage_bio_submit_read(args
.bio
);
404 EXPORT_SYMBOL(mpage_read_folio
);
407 * Writing is not so simple.
409 * If the page has buffers then they will be used for obtaining the disk
410 * mapping. We only support pages which are fully mapped-and-dirty, with a
411 * special case for pages which are unmapped at the end: end-of-file.
413 * If the page has no buffers (preferred) then the page is mapped here.
415 * If all blocks are found to be contiguous then the page can go into the
416 * BIO. Otherwise fall back to the mapping's writepage().
418 * FIXME: This code wants an estimate of how many pages are still to be
419 * written, so it can intelligently allocate a suitably-sized BIO. For now,
420 * just allocate full-size (16-page) BIOs.
425 sector_t last_block_in_bio
;
426 get_block_t
*get_block
;
430 * We have our BIO, so we can now mark the buffers clean. Make
431 * sure to only clean buffers which we know we'll be writing.
433 static void clean_buffers(struct page
*page
, unsigned first_unmapped
)
435 unsigned buffer_counter
= 0;
436 struct buffer_head
*bh
, *head
;
437 if (!page_has_buffers(page
))
439 head
= page_buffers(page
);
443 if (buffer_counter
++ == first_unmapped
)
445 clear_buffer_dirty(bh
);
446 bh
= bh
->b_this_page
;
447 } while (bh
!= head
);
450 * we cannot drop the bh if the page is not uptodate or a concurrent
451 * read_folio would fail to serialize with the bh and it would read from
452 * disk before we reach the platter.
454 if (buffer_heads_over_limit
&& PageUptodate(page
))
455 try_to_free_buffers(page_folio(page
));
459 * For situations where we want to clean all buffers attached to a page.
460 * We don't need to calculate how many buffers are attached to the page,
461 * we just need to specify a number larger than the maximum number of buffers.
463 void clean_page_buffers(struct page
*page
)
465 clean_buffers(page
, ~0U);
468 static int __mpage_writepage(struct folio
*folio
, struct writeback_control
*wbc
,
471 struct mpage_data
*mpd
= data
;
472 struct bio
*bio
= mpd
->bio
;
473 struct address_space
*mapping
= folio
->mapping
;
474 struct inode
*inode
= mapping
->host
;
475 const unsigned blkbits
= inode
->i_blkbits
;
476 const unsigned blocks_per_page
= PAGE_SIZE
>> blkbits
;
478 sector_t block_in_file
;
479 sector_t blocks
[MAX_BUF_PER_PAGE
];
481 unsigned first_unmapped
= blocks_per_page
;
482 struct block_device
*bdev
= NULL
;
484 sector_t boundary_block
= 0;
485 struct block_device
*boundary_bdev
= NULL
;
487 struct buffer_head map_bh
;
488 loff_t i_size
= i_size_read(inode
);
490 struct buffer_head
*head
= folio_buffers(folio
);
493 struct buffer_head
*bh
= head
;
495 /* If they're all mapped and dirty, do it */
498 BUG_ON(buffer_locked(bh
));
499 if (!buffer_mapped(bh
)) {
501 * unmapped dirty buffers are created by
502 * block_dirty_folio -> mmapped data
504 if (buffer_dirty(bh
))
506 if (first_unmapped
== blocks_per_page
)
507 first_unmapped
= page_block
;
511 if (first_unmapped
!= blocks_per_page
)
512 goto confused
; /* hole -> non-hole */
514 if (!buffer_dirty(bh
) || !buffer_uptodate(bh
))
517 if (bh
->b_blocknr
!= blocks
[page_block
-1] + 1)
520 blocks
[page_block
++] = bh
->b_blocknr
;
521 boundary
= buffer_boundary(bh
);
523 boundary_block
= bh
->b_blocknr
;
524 boundary_bdev
= bh
->b_bdev
;
527 } while ((bh
= bh
->b_this_page
) != head
);
533 * Page has buffers, but they are all unmapped. The page was
534 * created by pagein or read over a hole which was handled by
535 * block_read_full_folio(). If this address_space is also
536 * using mpage_readahead then this can rarely happen.
542 * The page has no buffers: map it to disk
544 BUG_ON(!folio_test_uptodate(folio
));
545 block_in_file
= (sector_t
)folio
->index
<< (PAGE_SHIFT
- blkbits
);
547 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
550 if (block_in_file
>= (i_size
+ (1 << blkbits
) - 1) >> blkbits
)
552 last_block
= (i_size
- 1) >> blkbits
;
553 map_bh
.b_folio
= folio
;
554 for (page_block
= 0; page_block
< blocks_per_page
; ) {
557 map_bh
.b_size
= 1 << blkbits
;
558 if (mpd
->get_block(inode
, block_in_file
, &map_bh
, 1))
560 if (!buffer_mapped(&map_bh
))
562 if (buffer_new(&map_bh
))
563 clean_bdev_bh_alias(&map_bh
);
564 if (buffer_boundary(&map_bh
)) {
565 boundary_block
= map_bh
.b_blocknr
;
566 boundary_bdev
= map_bh
.b_bdev
;
569 if (map_bh
.b_blocknr
!= blocks
[page_block
-1] + 1)
572 blocks
[page_block
++] = map_bh
.b_blocknr
;
573 boundary
= buffer_boundary(&map_bh
);
574 bdev
= map_bh
.b_bdev
;
575 if (block_in_file
== last_block
)
579 BUG_ON(page_block
== 0);
581 first_unmapped
= page_block
;
584 /* Don't bother writing beyond EOF, truncate will discard the folio */
585 if (folio_pos(folio
) >= i_size
)
587 length
= folio_size(folio
);
588 if (folio_pos(folio
) + length
> i_size
) {
590 * The page straddles i_size. It must be zeroed out on each
591 * and every writepage invocation because it may be mmapped.
592 * "A file is mapped in multiples of the page size. For a file
593 * that is not a multiple of the page size, the remaining memory
594 * is zeroed when mapped, and writes to that region are not
595 * written out to the file."
597 length
= i_size
- folio_pos(folio
);
598 folio_zero_segment(folio
, length
, folio_size(folio
));
602 * This page will go to BIO. Do we need to send this BIO off first?
604 if (bio
&& mpd
->last_block_in_bio
!= blocks
[0] - 1)
605 bio
= mpage_bio_submit_write(bio
);
609 bio
= bio_alloc(bdev
, BIO_MAX_VECS
,
610 REQ_OP_WRITE
| wbc_to_write_flags(wbc
),
612 bio
->bi_iter
.bi_sector
= blocks
[0] << (blkbits
- 9);
613 wbc_init_bio(wbc
, bio
);
617 * Must try to add the page before marking the buffer clean or
618 * the confused fail path above (OOM) will be very confused when
619 * it finds all bh marked clean (i.e. it will not write anything)
621 wbc_account_cgroup_owner(wbc
, &folio
->page
, folio_size(folio
));
622 length
= first_unmapped
<< blkbits
;
623 if (!bio_add_folio(bio
, folio
, length
, 0)) {
624 bio
= mpage_bio_submit_write(bio
);
628 clean_buffers(&folio
->page
, first_unmapped
);
630 BUG_ON(folio_test_writeback(folio
));
631 folio_start_writeback(folio
);
633 if (boundary
|| (first_unmapped
!= blocks_per_page
)) {
634 bio
= mpage_bio_submit_write(bio
);
635 if (boundary_block
) {
636 write_boundary_block(boundary_bdev
,
637 boundary_block
, 1 << blkbits
);
640 mpd
->last_block_in_bio
= blocks
[blocks_per_page
- 1];
646 bio
= mpage_bio_submit_write(bio
);
649 * The caller has a ref on the inode, so *mapping is stable
651 ret
= block_write_full_page(&folio
->page
, mpd
->get_block
, wbc
);
652 mapping_set_error(mapping
, ret
);
659 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
660 * @mapping: address space structure to write
661 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
662 * @get_block: the filesystem's block mapper function.
664 * This is a library function, which implements the writepages()
665 * address_space_operation.
668 mpage_writepages(struct address_space
*mapping
,
669 struct writeback_control
*wbc
, get_block_t get_block
)
671 struct mpage_data mpd
= {
672 .get_block
= get_block
,
674 struct blk_plug plug
;
677 blk_start_plug(&plug
);
678 ret
= write_cache_pages(mapping
, wbc
, __mpage_writepage
, &mpd
);
680 mpage_bio_submit_write(mpd
.bio
);
681 blk_finish_plug(&plug
);
684 EXPORT_SYMBOL(mpage_writepages
);