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 if (err)
53 folio_set_error(fi.folio);
54 else
55 folio_mark_uptodate(fi.folio);
56 folio_unlock(fi.folio);
57 }
58
59 bio_put(bio);
60 }
61
mpage_write_end_io(struct bio * bio)62 static void mpage_write_end_io(struct bio *bio)
63 {
64 struct folio_iter fi;
65 int err = blk_status_to_errno(bio->bi_status);
66
67 bio_for_each_folio_all(fi, bio) {
68 if (err) {
69 folio_set_error(fi.folio);
70 mapping_set_error(fi.folio->mapping, err);
71 }
72 folio_end_writeback(fi.folio);
73 }
74
75 bio_put(bio);
76 }
77
mpage_bio_submit_read(struct bio * bio)78 static struct bio *mpage_bio_submit_read(struct bio *bio)
79 {
80 bio->bi_end_io = mpage_read_end_io;
81 guard_bio_eod(bio);
82 submit_bio(bio);
83 return NULL;
84 }
85
mpage_bio_submit_write(struct bio * bio)86 static struct bio *mpage_bio_submit_write(struct bio *bio)
87 {
88 bio->bi_end_io = mpage_write_end_io;
89 guard_bio_eod(bio);
90 submit_bio(bio);
91 return NULL;
92 }
93
94 /*
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
98 * to get_block.
99 *
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.
103 */
map_buffer_to_folio(struct folio * folio,struct buffer_head * bh,int page_block)104 static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
105 int page_block)
106 {
107 struct inode *inode = folio->mapping->host;
108 struct buffer_head *page_bh, *head;
109 int block = 0;
110
111 head = folio_buffers(folio);
112 if (!head) {
113 /*
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
116 */
117 if (inode->i_blkbits == PAGE_SHIFT &&
118 buffer_uptodate(bh)) {
119 folio_mark_uptodate(folio);
120 return;
121 }
122 create_empty_buffers(&folio->page, i_blocksize(inode), 0);
123 head = folio_buffers(folio);
124 }
125
126 page_bh = head;
127 do {
128 if (block == page_block) {
129 page_bh->b_state = bh->b_state;
130 page_bh->b_bdev = bh->b_bdev;
131 page_bh->b_blocknr = bh->b_blocknr;
132 break;
133 }
134 page_bh = page_bh->b_this_page;
135 block++;
136 } while (page_bh != head);
137 }
138
139 struct mpage_readpage_args {
140 struct bio *bio;
141 struct folio *folio;
142 unsigned int nr_pages;
143 bool is_readahead;
144 sector_t last_block_in_bio;
145 struct buffer_head map_bh;
146 unsigned long first_logical_block;
147 get_block_t *get_block;
148 };
149
150 /*
151 * This is the worker routine which does all the work of mapping the disk
152 * blocks and constructs largest possible bios, submits them for IO if the
153 * blocks are not contiguous on the disk.
154 *
155 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
156 * represent the validity of its disk mapping and to decide when to do the next
157 * get_block() call.
158 */
do_mpage_readpage(struct mpage_readpage_args * args)159 static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
160 {
161 struct folio *folio = args->folio;
162 struct inode *inode = folio->mapping->host;
163 const unsigned blkbits = inode->i_blkbits;
164 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
165 const unsigned blocksize = 1 << blkbits;
166 struct buffer_head *map_bh = &args->map_bh;
167 sector_t block_in_file;
168 sector_t last_block;
169 sector_t last_block_in_file;
170 sector_t blocks[MAX_BUF_PER_PAGE];
171 unsigned page_block;
172 unsigned first_hole = blocks_per_page;
173 struct block_device *bdev = NULL;
174 int length;
175 int fully_mapped = 1;
176 blk_opf_t opf = REQ_OP_READ;
177 unsigned nblocks;
178 unsigned relative_block;
179 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
180
181 /* MAX_BUF_PER_PAGE, for example */
182 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
183
184 if (args->is_readahead) {
185 opf |= REQ_RAHEAD;
186 gfp |= __GFP_NORETRY | __GFP_NOWARN;
187 }
188
189 if (folio_buffers(folio))
190 goto confused;
191
192 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
193 last_block = block_in_file + args->nr_pages * blocks_per_page;
194 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
195 if (last_block > last_block_in_file)
196 last_block = last_block_in_file;
197 page_block = 0;
198
199 /*
200 * Map blocks using the result from the previous get_blocks call first.
201 */
202 nblocks = map_bh->b_size >> blkbits;
203 if (buffer_mapped(map_bh) &&
204 block_in_file > args->first_logical_block &&
205 block_in_file < (args->first_logical_block + nblocks)) {
206 unsigned map_offset = block_in_file - args->first_logical_block;
207 unsigned last = nblocks - map_offset;
208
209 for (relative_block = 0; ; relative_block++) {
210 if (relative_block == last) {
211 clear_buffer_mapped(map_bh);
212 break;
213 }
214 if (page_block == blocks_per_page)
215 break;
216 blocks[page_block] = map_bh->b_blocknr + map_offset +
217 relative_block;
218 page_block++;
219 block_in_file++;
220 }
221 bdev = map_bh->b_bdev;
222 }
223
224 /*
225 * Then do more get_blocks calls until we are done with this folio.
226 */
227 map_bh->b_folio = folio;
228 while (page_block < blocks_per_page) {
229 map_bh->b_state = 0;
230 map_bh->b_size = 0;
231
232 if (block_in_file < last_block) {
233 map_bh->b_size = (last_block-block_in_file) << blkbits;
234 if (args->get_block(inode, block_in_file, map_bh, 0))
235 goto confused;
236 args->first_logical_block = block_in_file;
237 }
238
239 if (!buffer_mapped(map_bh)) {
240 fully_mapped = 0;
241 if (first_hole == blocks_per_page)
242 first_hole = page_block;
243 page_block++;
244 block_in_file++;
245 continue;
246 }
247
248 /* some filesystems will copy data into the page during
249 * the get_block call, in which case we don't want to
250 * read it again. map_buffer_to_folio copies the data
251 * we just collected from get_block into the folio's buffers
252 * so read_folio doesn't have to repeat the get_block call
253 */
254 if (buffer_uptodate(map_bh)) {
255 map_buffer_to_folio(folio, map_bh, page_block);
256 goto confused;
257 }
258
259 if (first_hole != blocks_per_page)
260 goto confused; /* hole -> non-hole */
261
262 /* Contiguous blocks? */
263 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
264 goto confused;
265 nblocks = map_bh->b_size >> blkbits;
266 for (relative_block = 0; ; relative_block++) {
267 if (relative_block == nblocks) {
268 clear_buffer_mapped(map_bh);
269 break;
270 } else if (page_block == blocks_per_page)
271 break;
272 blocks[page_block] = map_bh->b_blocknr+relative_block;
273 page_block++;
274 block_in_file++;
275 }
276 bdev = map_bh->b_bdev;
277 }
278
279 if (first_hole != blocks_per_page) {
280 folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
281 if (first_hole == 0) {
282 folio_mark_uptodate(folio);
283 folio_unlock(folio);
284 goto out;
285 }
286 } else if (fully_mapped) {
287 folio_set_mappedtodisk(folio);
288 }
289
290 /*
291 * This folio will go to BIO. Do we need to send this BIO off first?
292 */
293 if (args->bio && (args->last_block_in_bio != blocks[0] - 1))
294 args->bio = mpage_bio_submit_read(args->bio);
295
296 alloc_new:
297 if (args->bio == NULL) {
298 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
299 gfp);
300 if (args->bio == NULL)
301 goto confused;
302 args->bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
303 }
304
305 length = first_hole << blkbits;
306 if (!bio_add_folio(args->bio, folio, length, 0)) {
307 args->bio = mpage_bio_submit_read(args->bio);
308 goto alloc_new;
309 }
310
311 relative_block = block_in_file - args->first_logical_block;
312 nblocks = map_bh->b_size >> blkbits;
313 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
314 (first_hole != blocks_per_page))
315 args->bio = mpage_bio_submit_read(args->bio);
316 else
317 args->last_block_in_bio = blocks[blocks_per_page - 1];
318 out:
319 return args->bio;
320
321 confused:
322 if (args->bio)
323 args->bio = mpage_bio_submit_read(args->bio);
324 if (!folio_test_uptodate(folio))
325 block_read_full_folio(folio, args->get_block);
326 else
327 folio_unlock(folio);
328 goto out;
329 }
330
331 /**
332 * mpage_readahead - start reads against pages
333 * @rac: Describes which pages to read.
334 * @get_block: The filesystem's block mapper function.
335 *
336 * This function walks the pages and the blocks within each page, building and
337 * emitting large BIOs.
338 *
339 * If anything unusual happens, such as:
340 *
341 * - encountering a page which has buffers
342 * - encountering a page which has a non-hole after a hole
343 * - encountering a page with non-contiguous blocks
344 *
345 * then this code just gives up and calls the buffer_head-based read function.
346 * It does handle a page which has holes at the end - that is a common case:
347 * the end-of-file on blocksize < PAGE_SIZE setups.
348 *
349 * BH_Boundary explanation:
350 *
351 * There is a problem. The mpage read code assembles several pages, gets all
352 * their disk mappings, and then submits them all. That's fine, but obtaining
353 * the disk mappings may require I/O. Reads of indirect blocks, for example.
354 *
355 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
356 * submitted in the following order:
357 *
358 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
359 *
360 * because the indirect block has to be read to get the mappings of blocks
361 * 13,14,15,16. Obviously, this impacts performance.
362 *
363 * So what we do it to allow the filesystem's get_block() function to set
364 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
365 * after this one will require I/O against a block which is probably close to
366 * this one. So you should push what I/O you have currently accumulated.
367 *
368 * This all causes the disk requests to be issued in the correct order.
369 */
mpage_readahead(struct readahead_control * rac,get_block_t get_block)370 void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
371 {
372 struct folio *folio;
373 struct mpage_readpage_args args = {
374 .get_block = get_block,
375 .is_readahead = true,
376 };
377
378 while ((folio = readahead_folio(rac))) {
379 prefetchw(&folio->flags);
380 args.folio = folio;
381 args.nr_pages = readahead_count(rac);
382 args.bio = do_mpage_readpage(&args);
383 }
384 if (args.bio)
385 mpage_bio_submit_read(args.bio);
386 }
387 EXPORT_SYMBOL(mpage_readahead);
388
389 /*
390 * This isn't called much at all
391 */
mpage_read_folio(struct folio * folio,get_block_t get_block)392 int mpage_read_folio(struct folio *folio, get_block_t get_block)
393 {
394 struct mpage_readpage_args args = {
395 .folio = folio,
396 .nr_pages = 1,
397 .get_block = get_block,
398 };
399
400 args.bio = do_mpage_readpage(&args);
401 if (args.bio)
402 mpage_bio_submit_read(args.bio);
403 return 0;
404 }
405 EXPORT_SYMBOL(mpage_read_folio);
406
407 /*
408 * Writing is not so simple.
409 *
410 * If the page has buffers then they will be used for obtaining the disk
411 * mapping. We only support pages which are fully mapped-and-dirty, with a
412 * special case for pages which are unmapped at the end: end-of-file.
413 *
414 * If the page has no buffers (preferred) then the page is mapped here.
415 *
416 * If all blocks are found to be contiguous then the page can go into the
417 * BIO. Otherwise fall back to the mapping's writepage().
418 *
419 * FIXME: This code wants an estimate of how many pages are still to be
420 * written, so it can intelligently allocate a suitably-sized BIO. For now,
421 * just allocate full-size (16-page) BIOs.
422 */
423
424 struct mpage_data {
425 struct bio *bio;
426 sector_t last_block_in_bio;
427 get_block_t *get_block;
428 };
429
430 /*
431 * We have our BIO, so we can now mark the buffers clean. Make
432 * sure to only clean buffers which we know we'll be writing.
433 */
clean_buffers(struct page * page,unsigned first_unmapped)434 static void clean_buffers(struct page *page, unsigned first_unmapped)
435 {
436 unsigned buffer_counter = 0;
437 struct buffer_head *bh, *head;
438 if (!page_has_buffers(page))
439 return;
440 head = page_buffers(page);
441 bh = head;
442
443 do {
444 if (buffer_counter++ == first_unmapped)
445 break;
446 clear_buffer_dirty(bh);
447 bh = bh->b_this_page;
448 } while (bh != head);
449
450 /*
451 * we cannot drop the bh if the page is not uptodate or a concurrent
452 * read_folio would fail to serialize with the bh and it would read from
453 * disk before we reach the platter.
454 */
455 if (buffer_heads_over_limit && PageUptodate(page))
456 try_to_free_buffers(page_folio(page));
457 }
458
459 /*
460 * For situations where we want to clean all buffers attached to a page.
461 * We don't need to calculate how many buffers are attached to the page,
462 * we just need to specify a number larger than the maximum number of buffers.
463 */
clean_page_buffers(struct page * page)464 void clean_page_buffers(struct page *page)
465 {
466 clean_buffers(page, ~0U);
467 }
468
__mpage_writepage(struct folio * folio,struct writeback_control * wbc,void * data)469 static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
470 void *data)
471 {
472 struct mpage_data *mpd = data;
473 struct bio *bio = mpd->bio;
474 struct address_space *mapping = folio->mapping;
475 struct inode *inode = mapping->host;
476 const unsigned blkbits = inode->i_blkbits;
477 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
478 sector_t last_block;
479 sector_t block_in_file;
480 sector_t blocks[MAX_BUF_PER_PAGE];
481 unsigned page_block;
482 unsigned first_unmapped = blocks_per_page;
483 struct block_device *bdev = NULL;
484 int boundary = 0;
485 sector_t boundary_block = 0;
486 struct block_device *boundary_bdev = NULL;
487 size_t length;
488 struct buffer_head map_bh;
489 loff_t i_size = i_size_read(inode);
490 int ret = 0;
491 struct buffer_head *head = folio_buffers(folio);
492
493 if (head) {
494 struct buffer_head *bh = head;
495
496 /* If they're all mapped and dirty, do it */
497 page_block = 0;
498 do {
499 BUG_ON(buffer_locked(bh));
500 if (!buffer_mapped(bh)) {
501 /*
502 * unmapped dirty buffers are created by
503 * block_dirty_folio -> mmapped data
504 */
505 if (buffer_dirty(bh))
506 goto confused;
507 if (first_unmapped == blocks_per_page)
508 first_unmapped = page_block;
509 continue;
510 }
511
512 if (first_unmapped != blocks_per_page)
513 goto confused; /* hole -> non-hole */
514
515 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
516 goto confused;
517 if (page_block) {
518 if (bh->b_blocknr != blocks[page_block-1] + 1)
519 goto confused;
520 }
521 blocks[page_block++] = bh->b_blocknr;
522 boundary = buffer_boundary(bh);
523 if (boundary) {
524 boundary_block = bh->b_blocknr;
525 boundary_bdev = bh->b_bdev;
526 }
527 bdev = bh->b_bdev;
528 } while ((bh = bh->b_this_page) != head);
529
530 if (first_unmapped)
531 goto page_is_mapped;
532
533 /*
534 * Page has buffers, but they are all unmapped. The page was
535 * created by pagein or read over a hole which was handled by
536 * block_read_full_folio(). If this address_space is also
537 * using mpage_readahead then this can rarely happen.
538 */
539 goto confused;
540 }
541
542 /*
543 * The page has no buffers: map it to disk
544 */
545 BUG_ON(!folio_test_uptodate(folio));
546 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
547 /*
548 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
549 * space.
550 */
551 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
552 goto page_is_mapped;
553 last_block = (i_size - 1) >> blkbits;
554 map_bh.b_folio = folio;
555 for (page_block = 0; page_block < blocks_per_page; ) {
556
557 map_bh.b_state = 0;
558 map_bh.b_size = 1 << blkbits;
559 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
560 goto confused;
561 if (!buffer_mapped(&map_bh))
562 goto confused;
563 if (buffer_new(&map_bh))
564 clean_bdev_bh_alias(&map_bh);
565 if (buffer_boundary(&map_bh)) {
566 boundary_block = map_bh.b_blocknr;
567 boundary_bdev = map_bh.b_bdev;
568 }
569 if (page_block) {
570 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
571 goto confused;
572 }
573 blocks[page_block++] = map_bh.b_blocknr;
574 boundary = buffer_boundary(&map_bh);
575 bdev = map_bh.b_bdev;
576 if (block_in_file == last_block)
577 break;
578 block_in_file++;
579 }
580 BUG_ON(page_block == 0);
581
582 first_unmapped = page_block;
583
584 page_is_mapped:
585 /* Don't bother writing beyond EOF, truncate will discard the folio */
586 if (folio_pos(folio) >= i_size)
587 goto confused;
588 length = folio_size(folio);
589 if (folio_pos(folio) + length > i_size) {
590 /*
591 * The page straddles i_size. It must be zeroed out on each
592 * and every writepage invocation because it may be mmapped.
593 * "A file is mapped in multiples of the page size. For a file
594 * that is not a multiple of the page size, the remaining memory
595 * is zeroed when mapped, and writes to that region are not
596 * written out to the file."
597 */
598 length = i_size - folio_pos(folio);
599 folio_zero_segment(folio, length, folio_size(folio));
600 }
601
602 /*
603 * This page will go to BIO. Do we need to send this BIO off first?
604 */
605 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
606 bio = mpage_bio_submit_write(bio);
607
608 alloc_new:
609 if (bio == NULL) {
610 bio = bio_alloc(bdev, BIO_MAX_VECS,
611 REQ_OP_WRITE | wbc_to_write_flags(wbc),
612 GFP_NOFS);
613 bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
614 wbc_init_bio(wbc, bio);
615 }
616
617 /*
618 * Must try to add the page before marking the buffer clean or
619 * the confused fail path above (OOM) will be very confused when
620 * it finds all bh marked clean (i.e. it will not write anything)
621 */
622 wbc_account_cgroup_owner(wbc, &folio->page, folio_size(folio));
623 length = first_unmapped << blkbits;
624 if (!bio_add_folio(bio, folio, length, 0)) {
625 bio = mpage_bio_submit_write(bio);
626 goto alloc_new;
627 }
628
629 clean_buffers(&folio->page, first_unmapped);
630
631 BUG_ON(folio_test_writeback(folio));
632 folio_start_writeback(folio);
633 folio_unlock(folio);
634 if (boundary || (first_unmapped != blocks_per_page)) {
635 bio = mpage_bio_submit_write(bio);
636 if (boundary_block) {
637 write_boundary_block(boundary_bdev,
638 boundary_block, 1 << blkbits);
639 }
640 } else {
641 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
642 }
643 goto out;
644
645 confused:
646 if (bio)
647 bio = mpage_bio_submit_write(bio);
648
649 /*
650 * The caller has a ref on the inode, so *mapping is stable
651 */
652 ret = block_write_full_page(&folio->page, mpd->get_block, wbc);
653 mapping_set_error(mapping, ret);
654 out:
655 mpd->bio = bio;
656 return ret;
657 }
658
659 /**
660 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
661 * @mapping: address space structure to write
662 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
663 * @get_block: the filesystem's block mapper function.
664 *
665 * This is a library function, which implements the writepages()
666 * address_space_operation.
667 */
668 int
mpage_writepages(struct address_space * mapping,struct writeback_control * wbc,get_block_t get_block)669 mpage_writepages(struct address_space *mapping,
670 struct writeback_control *wbc, get_block_t get_block)
671 {
672 struct mpage_data mpd = {
673 .get_block = get_block,
674 };
675 struct blk_plug plug;
676 int ret;
677
678 blk_start_plug(&plug);
679 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
680 if (mpd.bio)
681 mpage_bio_submit_write(mpd.bio);
682 blk_finish_plug(&plug);
683 return ret;
684 }
685 EXPORT_SYMBOL(mpage_writepages);
686
