1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * fs/mpage.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * Contains functions related to preparing and submitting BIOs which contain
8 * multiple pagecache pages.
9 *
10 * 15May2002 Andrew Morton
11 * Initial version
12 * 27Jun2002 axboe@suse.de
13 * use bio_add_page() to build bio's just the right size
14 */
15
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/mm.h>
19 #include <linux/kdev_t.h>
20 #include <linux/gfp.h>
21 #include <linux/bio.h>
22 #include <linux/fs.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>
32 #include "internal.h"
33
34 /*
35 * I/O completion handler for multipage BIOs.
36 *
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().
40 *
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.
45 */
46 static void mpage_read_end_io(struct bio *bio)
47 {
48 struct folio_iter fi;
49 int err = blk_status_to_errno(bio->bi_status);
50
51 bio_for_each_folio_all(fi, bio)
52 folio_end_read(fi.folio, err == 0);
53
54 bio_put(bio);
55 }
56
57 static void mpage_write_end_io(struct bio *bio)
58 {
59 struct folio_iter fi;
60 int err = blk_status_to_errno(bio->bi_status);
61
62 bio_for_each_folio_all(fi, bio) {
63 if (err)
64 mapping_set_error(fi.folio->mapping, err);
65 folio_end_writeback(fi.folio);
66 }
67
68 bio_put(bio);
69 }
70
71 static struct bio *mpage_bio_submit_read(struct bio *bio)
72 {
73 bio->bi_end_io = mpage_read_end_io;
74 guard_bio_eod(bio);
75 submit_bio(bio);
76 return NULL;
77 }
78
79 static struct bio *mpage_bio_submit_write(struct bio *bio)
80 {
81 bio->bi_end_io = mpage_write_end_io;
82 guard_bio_eod(bio);
83 submit_bio(bio);
84 return NULL;
85 }
86
87 /*
88 * support function for mpage_readahead. The fs supplied get_block might
89 * return an up to date buffer. This is used to map that buffer into
90 * the page, which allows read_folio to avoid triggering a duplicate call
91 * to get_block.
92 *
93 * The idea is to avoid adding buffers to pages that don't already have
94 * them. So when the buffer is up to date and the page size == block size,
95 * this marks the page up to date instead of adding new buffers.
96 */
97 static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
98 int page_block)
99 {
100 struct inode *inode = folio->mapping->host;
101 struct buffer_head *page_bh, *head;
102 int block = 0;
103
104 head = folio_buffers(folio);
105 if (!head) {
106 /*
107 * don't make any buffers if there is only one buffer on
108 * the folio and the folio just needs to be set up to date
109 */
110 if (inode->i_blkbits == PAGE_SHIFT &&
111 buffer_uptodate(bh)) {
112 folio_mark_uptodate(folio);
113 return;
114 }
115 head = create_empty_buffers(folio, i_blocksize(inode), 0);
116 }
117
118 page_bh = head;
119 do {
120 if (block == page_block) {
121 page_bh->b_state = bh->b_state;
122 page_bh->b_bdev = bh->b_bdev;
123 page_bh->b_blocknr = bh->b_blocknr;
124 break;
125 }
126 page_bh = page_bh->b_this_page;
127 block++;
128 } while (page_bh != head);
129 }
130
131 struct mpage_readpage_args {
132 struct bio *bio;
133 struct folio *folio;
134 unsigned int nr_pages;
135 bool is_readahead;
136 sector_t last_block_in_bio;
137 struct buffer_head map_bh;
138 unsigned long first_logical_block;
139 get_block_t *get_block;
140 };
141
142 /*
143 * This is the worker routine which does all the work of mapping the disk
144 * blocks and constructs largest possible bios, submits them for IO if the
145 * blocks are not contiguous on the disk.
146 *
147 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
148 * represent the validity of its disk mapping and to decide when to do the next
149 * get_block() call.
150 */
151 static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
152 {
153 struct folio *folio = args->folio;
154 struct inode *inode = folio->mapping->host;
155 const unsigned blkbits = inode->i_blkbits;
156 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
157 const unsigned blocksize = 1 << blkbits;
158 struct buffer_head *map_bh = &args->map_bh;
159 sector_t block_in_file;
160 sector_t last_block;
161 sector_t last_block_in_file;
162 sector_t first_block;
163 unsigned page_block;
164 unsigned first_hole = blocks_per_page;
165 struct block_device *bdev = NULL;
166 int length;
167 int fully_mapped = 1;
168 blk_opf_t opf = REQ_OP_READ;
169 unsigned nblocks;
170 unsigned relative_block;
171 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
172
173 /* MAX_BUF_PER_PAGE, for example */
174 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
175
176 if (args->is_readahead) {
177 opf |= REQ_RAHEAD;
178 gfp |= __GFP_NORETRY | __GFP_NOWARN;
179 }
180
181 if (folio_buffers(folio))
182 goto confused;
183
184 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
185 last_block = block_in_file + args->nr_pages * blocks_per_page;
186 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
187 if (last_block > last_block_in_file)
188 last_block = last_block_in_file;
189 page_block = 0;
190
191 /*
192 * Map blocks using the result from the previous get_blocks call first.
193 */
194 nblocks = map_bh->b_size >> blkbits;
195 if (buffer_mapped(map_bh) &&
196 block_in_file > args->first_logical_block &&
197 block_in_file < (args->first_logical_block + nblocks)) {
198 unsigned map_offset = block_in_file - args->first_logical_block;
199 unsigned last = nblocks - map_offset;
200
201 first_block = map_bh->b_blocknr + map_offset;
202 for (relative_block = 0; ; relative_block++) {
203 if (relative_block == last) {
204 clear_buffer_mapped(map_bh);
205 break;
206 }
207 if (page_block == blocks_per_page)
208 break;
209 page_block++;
210 block_in_file++;
211 }
212 bdev = map_bh->b_bdev;
213 }
214
215 /*
216 * Then do more get_blocks calls until we are done with this folio.
217 */
218 map_bh->b_folio = folio;
219 while (page_block < blocks_per_page) {
220 map_bh->b_state = 0;
221 map_bh->b_size = 0;
222
223 if (block_in_file < last_block) {
224 map_bh->b_size = (last_block-block_in_file) << blkbits;
225 if (args->get_block(inode, block_in_file, map_bh, 0))
226 goto confused;
227 args->first_logical_block = block_in_file;
228 }
229
230 if (!buffer_mapped(map_bh)) {
231 fully_mapped = 0;
232 if (first_hole == blocks_per_page)
233 first_hole = page_block;
234 page_block++;
235 block_in_file++;
236 continue;
237 }
238
239 /* some filesystems will copy data into the page during
240 * the get_block call, in which case we don't want to
241 * read it again. map_buffer_to_folio copies the data
242 * we just collected from get_block into the folio's buffers
243 * so read_folio doesn't have to repeat the get_block call
244 */
245 if (buffer_uptodate(map_bh)) {
246 map_buffer_to_folio(folio, map_bh, page_block);
247 goto confused;
248 }
249
250 if (first_hole != blocks_per_page)
251 goto confused; /* hole -> non-hole */
252
253 /* Contiguous blocks? */
254 if (!page_block)
255 first_block = map_bh->b_blocknr;
256 else if (first_block + page_block != map_bh->b_blocknr)
257 goto confused;
258 nblocks = map_bh->b_size >> blkbits;
259 for (relative_block = 0; ; relative_block++) {
260 if (relative_block == nblocks) {
261 clear_buffer_mapped(map_bh);
262 break;
263 } else if (page_block == blocks_per_page)
264 break;
265 page_block++;
266 block_in_file++;
267 }
268 bdev = map_bh->b_bdev;
269 }
270
271 if (first_hole != blocks_per_page) {
272 folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
273 if (first_hole == 0) {
274 folio_mark_uptodate(folio);
275 folio_unlock(folio);
276 goto out;
277 }
278 } else if (fully_mapped) {
279 folio_set_mappedtodisk(folio);
280 }
281
282 /*
283 * This folio will go to BIO. Do we need to send this BIO off first?
284 */
285 if (args->bio && (args->last_block_in_bio != first_block - 1))
286 args->bio = mpage_bio_submit_read(args->bio);
287
288 alloc_new:
289 if (args->bio == NULL) {
290 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
291 gfp);
292 if (args->bio == NULL)
293 goto confused;
294 args->bio->bi_iter.bi_sector = first_block << (blkbits - 9);
295 }
296
297 length = first_hole << blkbits;
298 if (!bio_add_folio(args->bio, folio, length, 0)) {
299 args->bio = mpage_bio_submit_read(args->bio);
300 goto alloc_new;
301 }
302
303 relative_block = block_in_file - args->first_logical_block;
304 nblocks = map_bh->b_size >> blkbits;
305 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
306 (first_hole != blocks_per_page))
307 args->bio = mpage_bio_submit_read(args->bio);
308 else
309 args->last_block_in_bio = first_block + blocks_per_page - 1;
310 out:
311 return args->bio;
312
313 confused:
314 if (args->bio)
315 args->bio = mpage_bio_submit_read(args->bio);
316 if (!folio_test_uptodate(folio))
317 block_read_full_folio(folio, args->get_block);
318 else
319 folio_unlock(folio);
320 goto out;
321 }
322
323 /**
324 * mpage_readahead - start reads against pages
325 * @rac: Describes which pages to read.
326 * @get_block: The filesystem's block mapper function.
327 *
328 * This function walks the pages and the blocks within each page, building and
329 * emitting large BIOs.
330 *
331 * If anything unusual happens, such as:
332 *
333 * - encountering a page which has buffers
334 * - encountering a page which has a non-hole after a hole
335 * - encountering a page with non-contiguous blocks
336 *
337 * then this code just gives up and calls the buffer_head-based read function.
338 * It does handle a page which has holes at the end - that is a common case:
339 * the end-of-file on blocksize < PAGE_SIZE setups.
340 *
341 * BH_Boundary explanation:
342 *
343 * There is a problem. The mpage read code assembles several pages, gets all
344 * their disk mappings, and then submits them all. That's fine, but obtaining
345 * the disk mappings may require I/O. Reads of indirect blocks, for example.
346 *
347 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
348 * submitted in the following order:
349 *
350 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
351 *
352 * because the indirect block has to be read to get the mappings of blocks
353 * 13,14,15,16. Obviously, this impacts performance.
354 *
355 * So what we do it to allow the filesystem's get_block() function to set
356 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
357 * after this one will require I/O against a block which is probably close to
358 * this one. So you should push what I/O you have currently accumulated.
359 *
360 * This all causes the disk requests to be issued in the correct order.
361 */
362 void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
363 {
364 struct folio *folio;
365 struct mpage_readpage_args args = {
366 .get_block = get_block,
367 .is_readahead = true,
368 };
369
370 while ((folio = readahead_folio(rac))) {
371 prefetchw(&folio->flags);
372 args.folio = folio;
373 args.nr_pages = readahead_count(rac);
374 args.bio = do_mpage_readpage(&args);
375 }
376 if (args.bio)
377 mpage_bio_submit_read(args.bio);
378 }
379 EXPORT_SYMBOL(mpage_readahead);
380
381 /*
382 * This isn't called much at all
383 */
384 int mpage_read_folio(struct folio *folio, get_block_t get_block)
385 {
386 struct mpage_readpage_args args = {
387 .folio = folio,
388 .nr_pages = 1,
389 .get_block = get_block,
390 };
391
392 args.bio = do_mpage_readpage(&args);
393 if (args.bio)
394 mpage_bio_submit_read(args.bio);
395 return 0;
396 }
397 EXPORT_SYMBOL(mpage_read_folio);
398
399 /*
400 * Writing is not so simple.
401 *
402 * If the page has buffers then they will be used for obtaining the disk
403 * mapping. We only support pages which are fully mapped-and-dirty, with a
404 * special case for pages which are unmapped at the end: end-of-file.
405 *
406 * If the page has no buffers (preferred) then the page is mapped here.
407 *
408 * If all blocks are found to be contiguous then the page can go into the
409 * BIO. Otherwise fall back to the mapping's writepage().
410 *
411 * FIXME: This code wants an estimate of how many pages are still to be
412 * written, so it can intelligently allocate a suitably-sized BIO. For now,
413 * just allocate full-size (16-page) BIOs.
414 */
415
416 struct mpage_data {
417 struct bio *bio;
418 sector_t last_block_in_bio;
419 get_block_t *get_block;
420 };
421
422 /*
423 * We have our BIO, so we can now mark the buffers clean. Make
424 * sure to only clean buffers which we know we'll be writing.
425 */
426 static void clean_buffers(struct folio *folio, unsigned first_unmapped)
427 {
428 unsigned buffer_counter = 0;
429 struct buffer_head *bh, *head = folio_buffers(folio);
430
431 if (!head)
432 return;
433 bh = head;
434
435 do {
436 if (buffer_counter++ == first_unmapped)
437 break;
438 clear_buffer_dirty(bh);
439 bh = bh->b_this_page;
440 } while (bh != head);
441
442 /*
443 * we cannot drop the bh if the page is not uptodate or a concurrent
444 * read_folio would fail to serialize with the bh and it would read from
445 * disk before we reach the platter.
446 */
447 if (buffer_heads_over_limit && folio_test_uptodate(folio))
448 try_to_free_buffers(folio);
449 }
450
451 static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
452 void *data)
453 {
454 struct mpage_data *mpd = data;
455 struct bio *bio = mpd->bio;
456 struct address_space *mapping = folio->mapping;
457 struct inode *inode = mapping->host;
458 const unsigned blkbits = inode->i_blkbits;
459 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
460 sector_t last_block;
461 sector_t block_in_file;
462 sector_t first_block;
463 unsigned page_block;
464 unsigned first_unmapped = blocks_per_page;
465 struct block_device *bdev = NULL;
466 int boundary = 0;
467 sector_t boundary_block = 0;
468 struct block_device *boundary_bdev = NULL;
469 size_t length;
470 struct buffer_head map_bh;
471 loff_t i_size = i_size_read(inode);
472 int ret = 0;
473 struct buffer_head *head = folio_buffers(folio);
474
475 if (head) {
476 struct buffer_head *bh = head;
477
478 /* If they're all mapped and dirty, do it */
479 page_block = 0;
480 do {
481 BUG_ON(buffer_locked(bh));
482 if (!buffer_mapped(bh)) {
483 /*
484 * unmapped dirty buffers are created by
485 * block_dirty_folio -> mmapped data
486 */
487 if (buffer_dirty(bh))
488 goto confused;
489 if (first_unmapped == blocks_per_page)
490 first_unmapped = page_block;
491 continue;
492 }
493
494 if (first_unmapped != blocks_per_page)
495 goto confused; /* hole -> non-hole */
496
497 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
498 goto confused;
499 if (page_block) {
500 if (bh->b_blocknr != first_block + page_block)
501 goto confused;
502 } else {
503 first_block = bh->b_blocknr;
504 }
505 page_block++;
506 boundary = buffer_boundary(bh);
507 if (boundary) {
508 boundary_block = bh->b_blocknr;
509 boundary_bdev = bh->b_bdev;
510 }
511 bdev = bh->b_bdev;
512 } while ((bh = bh->b_this_page) != head);
513
514 if (first_unmapped)
515 goto page_is_mapped;
516
517 /*
518 * Page has buffers, but they are all unmapped. The page was
519 * created by pagein or read over a hole which was handled by
520 * block_read_full_folio(). If this address_space is also
521 * using mpage_readahead then this can rarely happen.
522 */
523 goto confused;
524 }
525
526 /*
527 * The page has no buffers: map it to disk
528 */
529 BUG_ON(!folio_test_uptodate(folio));
530 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
531 /*
532 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
533 * space.
534 */
535 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
536 goto page_is_mapped;
537 last_block = (i_size - 1) >> blkbits;
538 map_bh.b_folio = folio;
539 for (page_block = 0; page_block < blocks_per_page; ) {
540
541 map_bh.b_state = 0;
542 map_bh.b_size = 1 << blkbits;
543 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
544 goto confused;
545 if (!buffer_mapped(&map_bh))
546 goto confused;
547 if (buffer_new(&map_bh))
548 clean_bdev_bh_alias(&map_bh);
549 if (buffer_boundary(&map_bh)) {
550 boundary_block = map_bh.b_blocknr;
551 boundary_bdev = map_bh.b_bdev;
552 }
553 if (page_block) {
554 if (map_bh.b_blocknr != first_block + page_block)
555 goto confused;
556 } else {
557 first_block = map_bh.b_blocknr;
558 }
559 page_block++;
560 boundary = buffer_boundary(&map_bh);
561 bdev = map_bh.b_bdev;
562 if (block_in_file == last_block)
563 break;
564 block_in_file++;
565 }
566 BUG_ON(page_block == 0);
567
568 first_unmapped = page_block;
569
570 page_is_mapped:
571 /* Don't bother writing beyond EOF, truncate will discard the folio */
572 if (folio_pos(folio) >= i_size)
573 goto confused;
574 length = folio_size(folio);
575 if (folio_pos(folio) + length > i_size) {
576 /*
577 * The page straddles i_size. It must be zeroed out on each
578 * and every writepage invocation because it may be mmapped.
579 * "A file is mapped in multiples of the page size. For a file
580 * that is not a multiple of the page size, the remaining memory
581 * is zeroed when mapped, and writes to that region are not
582 * written out to the file."
583 */
584 length = i_size - folio_pos(folio);
585 folio_zero_segment(folio, length, folio_size(folio));
586 }
587
588 /*
589 * This page will go to BIO. Do we need to send this BIO off first?
590 */
591 if (bio && mpd->last_block_in_bio != first_block - 1)
592 bio = mpage_bio_submit_write(bio);
593
594 alloc_new:
595 if (bio == NULL) {
596 bio = bio_alloc(bdev, BIO_MAX_VECS,
597 REQ_OP_WRITE | wbc_to_write_flags(wbc),
598 GFP_NOFS);
599 bio->bi_iter.bi_sector = first_block << (blkbits - 9);
600 wbc_init_bio(wbc, bio);
601 bio->bi_write_hint = inode->i_write_hint;
602 }
603
604 /*
605 * Must try to add the page before marking the buffer clean or
606 * the confused fail path above (OOM) will be very confused when
607 * it finds all bh marked clean (i.e. it will not write anything)
608 */
609 wbc_account_cgroup_owner(wbc, &folio->page, folio_size(folio));
610 length = first_unmapped << blkbits;
611 if (!bio_add_folio(bio, folio, length, 0)) {
612 bio = mpage_bio_submit_write(bio);
613 goto alloc_new;
614 }
615
616 clean_buffers(folio, first_unmapped);
617
618 BUG_ON(folio_test_writeback(folio));
619 folio_start_writeback(folio);
620 folio_unlock(folio);
621 if (boundary || (first_unmapped != blocks_per_page)) {
622 bio = mpage_bio_submit_write(bio);
623 if (boundary_block) {
624 write_boundary_block(boundary_bdev,
625 boundary_block, 1 << blkbits);
626 }
627 } else {
628 mpd->last_block_in_bio = first_block + blocks_per_page - 1;
629 }
630 goto out;
631
632 confused:
633 if (bio)
634 bio = mpage_bio_submit_write(bio);
635
636 /*
637 * The caller has a ref on the inode, so *mapping is stable
638 */
639 ret = block_write_full_folio(folio, wbc, mpd->get_block);
640 mapping_set_error(mapping, ret);
641 out:
642 mpd->bio = bio;
643 return ret;
644 }
645
646 /**
647 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
648 * @mapping: address space structure to write
649 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
650 * @get_block: the filesystem's block mapper function.
651 *
652 * This is a library function, which implements the writepages()
653 * address_space_operation.
654 */
655 int
656 mpage_writepages(struct address_space *mapping,
657 struct writeback_control *wbc, get_block_t get_block)
658 {
659 struct mpage_data mpd = {
660 .get_block = get_block,
661 };
662 struct blk_plug plug;
663 int ret;
664
665 blk_start_plug(&plug);
666 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
667 if (mpd.bio)
668 mpage_bio_submit_write(mpd.bio);
669 blk_finish_plug(&plug);
670 return ret;
671 }
672 EXPORT_SYMBOL(mpage_writepages);
673
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.