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 */
mpage_read_end_io(struct bio * bio)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
mpage_write_end_io(struct bio * bio)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
mpage_bio_submit_read(struct bio * bio)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
mpage_bio_submit_write(struct bio * bio)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 */
map_buffer_to_folio(struct folio * folio,struct buffer_head * bh,int page_block)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 == folio_shift(folio) &&
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 */
do_mpage_readpage(struct mpage_readpage_args * args)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_folio = folio_size(folio) >> 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_folio;
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 if (args->is_readahead) {
174 opf |= REQ_RAHEAD;
175 gfp |= __GFP_NORETRY | __GFP_NOWARN;
176 }
177
178 if (folio_buffers(folio))
179 goto confused;
180
181 block_in_file = folio_pos(folio) >> blkbits;
182 last_block = block_in_file + ((args->nr_pages * PAGE_SIZE) >> blkbits);
183 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
184 if (last_block > last_block_in_file)
185 last_block = last_block_in_file;
186 page_block = 0;
187
188 /*
189 * Map blocks using the result from the previous get_blocks call first.
190 */
191 nblocks = map_bh->b_size >> blkbits;
192 if (buffer_mapped(map_bh) &&
193 block_in_file > args->first_logical_block &&
194 block_in_file < (args->first_logical_block + nblocks)) {
195 unsigned map_offset = block_in_file - args->first_logical_block;
196 unsigned last = nblocks - map_offset;
197
198 first_block = map_bh->b_blocknr + map_offset;
199 for (relative_block = 0; ; relative_block++) {
200 if (relative_block == last) {
201 clear_buffer_mapped(map_bh);
202 break;
203 }
204 if (page_block == blocks_per_folio)
205 break;
206 page_block++;
207 block_in_file++;
208 }
209 bdev = map_bh->b_bdev;
210 }
211
212 /*
213 * Then do more get_blocks calls until we are done with this folio.
214 */
215 map_bh->b_folio = folio;
216 while (page_block < blocks_per_folio) {
217 map_bh->b_state = 0;
218 map_bh->b_size = 0;
219
220 if (block_in_file < last_block) {
221 map_bh->b_size = (last_block-block_in_file) << blkbits;
222 if (args->get_block(inode, block_in_file, map_bh, 0))
223 goto confused;
224 args->first_logical_block = block_in_file;
225 }
226
227 if (!buffer_mapped(map_bh)) {
228 fully_mapped = 0;
229 if (first_hole == blocks_per_folio)
230 first_hole = page_block;
231 page_block++;
232 block_in_file++;
233 continue;
234 }
235
236 /* some filesystems will copy data into the page during
237 * the get_block call, in which case we don't want to
238 * read it again. map_buffer_to_folio copies the data
239 * we just collected from get_block into the folio's buffers
240 * so read_folio doesn't have to repeat the get_block call
241 */
242 if (buffer_uptodate(map_bh)) {
243 map_buffer_to_folio(folio, map_bh, page_block);
244 goto confused;
245 }
246
247 if (first_hole != blocks_per_folio)
248 goto confused; /* hole -> non-hole */
249
250 /* Contiguous blocks? */
251 if (!page_block)
252 first_block = map_bh->b_blocknr;
253 else if (first_block + page_block != map_bh->b_blocknr)
254 goto confused;
255 nblocks = map_bh->b_size >> blkbits;
256 for (relative_block = 0; ; relative_block++) {
257 if (relative_block == nblocks) {
258 clear_buffer_mapped(map_bh);
259 break;
260 } else if (page_block == blocks_per_folio)
261 break;
262 page_block++;
263 block_in_file++;
264 }
265 bdev = map_bh->b_bdev;
266 }
267
268 if (first_hole != blocks_per_folio) {
269 folio_zero_segment(folio, first_hole << blkbits, folio_size(folio));
270 if (first_hole == 0) {
271 folio_mark_uptodate(folio);
272 folio_unlock(folio);
273 goto out;
274 }
275 } else if (fully_mapped) {
276 folio_set_mappedtodisk(folio);
277 }
278
279 /*
280 * This folio will go to BIO. Do we need to send this BIO off first?
281 */
282 if (args->bio && (args->last_block_in_bio != first_block - 1))
283 args->bio = mpage_bio_submit_read(args->bio);
284
285 alloc_new:
286 if (args->bio == NULL) {
287 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
288 gfp);
289 if (args->bio == NULL)
290 goto confused;
291 args->bio->bi_iter.bi_sector = first_block << (blkbits - 9);
292 }
293
294 length = first_hole << blkbits;
295 if (!bio_add_folio(args->bio, folio, length, 0)) {
296 args->bio = mpage_bio_submit_read(args->bio);
297 goto alloc_new;
298 }
299
300 relative_block = block_in_file - args->first_logical_block;
301 nblocks = map_bh->b_size >> blkbits;
302 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
303 (first_hole != blocks_per_folio))
304 args->bio = mpage_bio_submit_read(args->bio);
305 else
306 args->last_block_in_bio = first_block + blocks_per_folio - 1;
307 out:
308 return args->bio;
309
310 confused:
311 if (args->bio)
312 args->bio = mpage_bio_submit_read(args->bio);
313 if (!folio_test_uptodate(folio))
314 block_read_full_folio(folio, args->get_block);
315 else
316 folio_unlock(folio);
317 goto out;
318 }
319
320 /**
321 * mpage_readahead - start reads against pages
322 * @rac: Describes which pages to read.
323 * @get_block: The filesystem's block mapper function.
324 *
325 * This function walks the pages and the blocks within each page, building and
326 * emitting large BIOs.
327 *
328 * If anything unusual happens, such as:
329 *
330 * - encountering a page which has buffers
331 * - encountering a page which has a non-hole after a hole
332 * - encountering a page with non-contiguous blocks
333 *
334 * then this code just gives up and calls the buffer_head-based read function.
335 * It does handle a page which has holes at the end - that is a common case:
336 * the end-of-file on blocksize < PAGE_SIZE setups.
337 *
338 * BH_Boundary explanation:
339 *
340 * There is a problem. The mpage read code assembles several pages, gets all
341 * their disk mappings, and then submits them all. That's fine, but obtaining
342 * the disk mappings may require I/O. Reads of indirect blocks, for example.
343 *
344 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
345 * submitted in the following order:
346 *
347 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
348 *
349 * because the indirect block has to be read to get the mappings of blocks
350 * 13,14,15,16. Obviously, this impacts performance.
351 *
352 * So what we do it to allow the filesystem's get_block() function to set
353 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
354 * after this one will require I/O against a block which is probably close to
355 * this one. So you should push what I/O you have currently accumulated.
356 *
357 * This all causes the disk requests to be issued in the correct order.
358 */
mpage_readahead(struct readahead_control * rac,get_block_t get_block)359 void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
360 {
361 struct folio *folio;
362 struct mpage_readpage_args args = {
363 .get_block = get_block,
364 .is_readahead = true,
365 };
366
367 while ((folio = readahead_folio(rac))) {
368 prefetchw(&folio->flags);
369 args.folio = folio;
370 args.nr_pages = readahead_count(rac);
371 args.bio = do_mpage_readpage(&args);
372 }
373 if (args.bio)
374 mpage_bio_submit_read(args.bio);
375 }
376 EXPORT_SYMBOL(mpage_readahead);
377
378 /*
379 * This isn't called much at all
380 */
mpage_read_folio(struct folio * folio,get_block_t get_block)381 int mpage_read_folio(struct folio *folio, get_block_t get_block)
382 {
383 struct mpage_readpage_args args = {
384 .folio = folio,
385 .nr_pages = folio_nr_pages(folio),
386 .get_block = get_block,
387 };
388
389 args.bio = do_mpage_readpage(&args);
390 if (args.bio)
391 mpage_bio_submit_read(args.bio);
392 return 0;
393 }
394 EXPORT_SYMBOL(mpage_read_folio);
395
396 /*
397 * Writing is not so simple.
398 *
399 * If the page has buffers then they will be used for obtaining the disk
400 * mapping. We only support pages which are fully mapped-and-dirty, with a
401 * special case for pages which are unmapped at the end: end-of-file.
402 *
403 * If the page has no buffers (preferred) then the page is mapped here.
404 *
405 * If all blocks are found to be contiguous then the page can go into the
406 * BIO. Otherwise fall back to the mapping's writepage().
407 *
408 * FIXME: This code wants an estimate of how many pages are still to be
409 * written, so it can intelligently allocate a suitably-sized BIO. For now,
410 * just allocate full-size (16-page) BIOs.
411 */
412
413 struct mpage_data {
414 struct bio *bio;
415 sector_t last_block_in_bio;
416 get_block_t *get_block;
417 };
418
419 /*
420 * We have our BIO, so we can now mark the buffers clean. Make
421 * sure to only clean buffers which we know we'll be writing.
422 */
clean_buffers(struct folio * folio,unsigned first_unmapped)423 static void clean_buffers(struct folio *folio, unsigned first_unmapped)
424 {
425 unsigned buffer_counter = 0;
426 struct buffer_head *bh, *head = folio_buffers(folio);
427
428 if (!head)
429 return;
430 bh = head;
431
432 do {
433 if (buffer_counter++ == first_unmapped)
434 break;
435 clear_buffer_dirty(bh);
436 bh = bh->b_this_page;
437 } while (bh != head);
438
439 /*
440 * we cannot drop the bh if the page is not uptodate or a concurrent
441 * read_folio would fail to serialize with the bh and it would read from
442 * disk before we reach the platter.
443 */
444 if (buffer_heads_over_limit && folio_test_uptodate(folio))
445 try_to_free_buffers(folio);
446 }
447
mpage_write_folio(struct writeback_control * wbc,struct folio * folio,struct mpage_data * mpd)448 static int mpage_write_folio(struct writeback_control *wbc, struct folio *folio,
449 struct mpage_data *mpd)
450 {
451 struct bio *bio = mpd->bio;
452 struct address_space *mapping = folio->mapping;
453 struct inode *inode = mapping->host;
454 const unsigned blkbits = inode->i_blkbits;
455 const unsigned blocks_per_folio = folio_size(folio) >> blkbits;
456 sector_t last_block;
457 sector_t block_in_file;
458 sector_t first_block;
459 unsigned page_block;
460 unsigned first_unmapped = blocks_per_folio;
461 struct block_device *bdev = NULL;
462 int boundary = 0;
463 sector_t boundary_block = 0;
464 struct block_device *boundary_bdev = NULL;
465 size_t length;
466 struct buffer_head map_bh;
467 loff_t i_size = i_size_read(inode);
468 int ret = 0;
469 struct buffer_head *head = folio_buffers(folio);
470
471 if (head) {
472 struct buffer_head *bh = head;
473
474 /* If they're all mapped and dirty, do it */
475 page_block = 0;
476 do {
477 BUG_ON(buffer_locked(bh));
478 if (!buffer_mapped(bh)) {
479 /*
480 * unmapped dirty buffers are created by
481 * block_dirty_folio -> mmapped data
482 */
483 if (buffer_dirty(bh))
484 goto confused;
485 if (first_unmapped == blocks_per_folio)
486 first_unmapped = page_block;
487 continue;
488 }
489
490 if (first_unmapped != blocks_per_folio)
491 goto confused; /* hole -> non-hole */
492
493 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
494 goto confused;
495 if (page_block) {
496 if (bh->b_blocknr != first_block + page_block)
497 goto confused;
498 } else {
499 first_block = bh->b_blocknr;
500 }
501 page_block++;
502 boundary = buffer_boundary(bh);
503 if (boundary) {
504 boundary_block = bh->b_blocknr;
505 boundary_bdev = bh->b_bdev;
506 }
507 bdev = bh->b_bdev;
508 } while ((bh = bh->b_this_page) != head);
509
510 if (first_unmapped)
511 goto page_is_mapped;
512
513 /*
514 * Page has buffers, but they are all unmapped. The page was
515 * created by pagein or read over a hole which was handled by
516 * block_read_full_folio(). If this address_space is also
517 * using mpage_readahead then this can rarely happen.
518 */
519 goto confused;
520 }
521
522 /*
523 * The page has no buffers: map it to disk
524 */
525 BUG_ON(!folio_test_uptodate(folio));
526 block_in_file = folio_pos(folio) >> blkbits;
527 /*
528 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
529 * space.
530 */
531 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
532 goto page_is_mapped;
533 last_block = (i_size - 1) >> blkbits;
534 map_bh.b_folio = folio;
535 for (page_block = 0; page_block < blocks_per_folio; ) {
536
537 map_bh.b_state = 0;
538 map_bh.b_size = 1 << blkbits;
539 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
540 goto confused;
541 if (!buffer_mapped(&map_bh))
542 goto confused;
543 if (buffer_new(&map_bh))
544 clean_bdev_bh_alias(&map_bh);
545 if (buffer_boundary(&map_bh)) {
546 boundary_block = map_bh.b_blocknr;
547 boundary_bdev = map_bh.b_bdev;
548 }
549 if (page_block) {
550 if (map_bh.b_blocknr != first_block + page_block)
551 goto confused;
552 } else {
553 first_block = map_bh.b_blocknr;
554 }
555 page_block++;
556 boundary = buffer_boundary(&map_bh);
557 bdev = map_bh.b_bdev;
558 if (block_in_file == last_block)
559 break;
560 block_in_file++;
561 }
562 BUG_ON(page_block == 0);
563
564 first_unmapped = page_block;
565
566 page_is_mapped:
567 /* Don't bother writing beyond EOF, truncate will discard the folio */
568 if (folio_pos(folio) >= i_size)
569 goto confused;
570 length = folio_size(folio);
571 if (folio_pos(folio) + length > i_size) {
572 /*
573 * The page straddles i_size. It must be zeroed out on each
574 * and every writepage invocation because it may be mmapped.
575 * "A file is mapped in multiples of the page size. For a file
576 * that is not a multiple of the page size, the remaining memory
577 * is zeroed when mapped, and writes to that region are not
578 * written out to the file."
579 */
580 length = i_size - folio_pos(folio);
581 folio_zero_segment(folio, length, folio_size(folio));
582 }
583
584 /*
585 * This page will go to BIO. Do we need to send this BIO off first?
586 */
587 if (bio && mpd->last_block_in_bio != first_block - 1)
588 bio = mpage_bio_submit_write(bio);
589
590 alloc_new:
591 if (bio == NULL) {
592 bio = bio_alloc(bdev, BIO_MAX_VECS,
593 REQ_OP_WRITE | wbc_to_write_flags(wbc),
594 GFP_NOFS);
595 bio->bi_iter.bi_sector = first_block << (blkbits - 9);
596 wbc_init_bio(wbc, bio);
597 bio->bi_write_hint = inode->i_write_hint;
598 }
599
600 /*
601 * Must try to add the page before marking the buffer clean or
602 * the confused fail path above (OOM) will be very confused when
603 * it finds all bh marked clean (i.e. it will not write anything)
604 */
605 wbc_account_cgroup_owner(wbc, folio, folio_size(folio));
606 length = first_unmapped << blkbits;
607 if (!bio_add_folio(bio, folio, length, 0)) {
608 bio = mpage_bio_submit_write(bio);
609 goto alloc_new;
610 }
611
612 clean_buffers(folio, first_unmapped);
613
614 BUG_ON(folio_test_writeback(folio));
615 folio_start_writeback(folio);
616 folio_unlock(folio);
617 if (boundary || (first_unmapped != blocks_per_folio)) {
618 bio = mpage_bio_submit_write(bio);
619 if (boundary_block) {
620 write_boundary_block(boundary_bdev,
621 boundary_block, 1 << blkbits);
622 }
623 } else {
624 mpd->last_block_in_bio = first_block + blocks_per_folio - 1;
625 }
626 goto out;
627
628 confused:
629 if (bio)
630 bio = mpage_bio_submit_write(bio);
631
632 /*
633 * The caller has a ref on the inode, so *mapping is stable
634 */
635 ret = block_write_full_folio(folio, wbc, mpd->get_block);
636 mapping_set_error(mapping, ret);
637 out:
638 mpd->bio = bio;
639 return ret;
640 }
641
642 /**
643 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
644 * @mapping: address space structure to write
645 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
646 * @get_block: the filesystem's block mapper function.
647 *
648 * This is a library function, which implements the writepages()
649 * address_space_operation.
650 */
651 int
mpage_writepages(struct address_space * mapping,struct writeback_control * wbc,get_block_t get_block)652 mpage_writepages(struct address_space *mapping,
653 struct writeback_control *wbc, get_block_t get_block)
654 {
655 struct mpage_data mpd = {
656 .get_block = get_block,
657 };
658 struct folio *folio = NULL;
659 struct blk_plug plug;
660 int error;
661
662 blk_start_plug(&plug);
663 while ((folio = writeback_iter(mapping, wbc, folio, &error)))
664 error = mpage_write_folio(wbc, folio, &mpd);
665 if (mpd.bio)
666 mpage_bio_submit_write(mpd.bio);
667 blk_finish_plug(&plug);
668 return error;
669 }
670 EXPORT_SYMBOL(mpage_writepages);
671