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 */
46static void mpage_read_end_io(struct bio *bio)
47{
48 struct folio_iter fi;
49 int err = blk_status_to_errno(status: bio->bi_status);
50
51 bio_for_each_folio_all(fi, bio)
52 folio_end_read(folio: fi.folio, success: err == 0);
53
54 bio_put(bio);
55}
56
57static void mpage_write_end_io(struct bio *bio)
58{
59 struct folio_iter fi;
60 int err = blk_status_to_errno(status: bio->bi_status);
61
62 bio_for_each_folio_all(fi, bio) {
63 if (err)
64 mapping_set_error(mapping: fi.folio->mapping, error: err);
65 folio_end_writeback(folio: fi.folio);
66 }
67
68 bio_put(bio);
69}
70
71static 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
79static 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 */
97static 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, blocksize: i_blocksize(node: inode), b_state: 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
131struct 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 */
151static void 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(mapping: 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(bh: 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(bh: 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(bh: 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(bh: map_bh)) {
243 map_buffer_to_folio(folio, bh: 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(bh: 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, start: first_hole << blkbits, xend: 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(bio: args->bio);
284
285alloc_new:
286 if (args->bio == NULL) {
287 args->bio = bio_alloc(bdev, nr_vecs: bio_max_segs(nr_segs: args->nr_pages), opf,
288 gfp_mask: 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(bio: args->bio, folio, len: length, off: 0)) {
296 args->bio = mpage_bio_submit_read(bio: 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(bh: map_bh) && relative_block == nblocks) ||
303 (first_hole != blocks_per_folio))
304 args->bio = mpage_bio_submit_read(bio: args->bio);
305 else
306 args->last_block_in_bio = first_block + blocks_per_folio - 1;
307out:
308 return;
309
310confused:
311 if (args->bio)
312 args->bio = mpage_bio_submit_read(bio: 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 */
359void 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(ractl: rac))) {
368 prefetchw(x: &folio->flags);
369 args.folio = folio;
370 args.nr_pages = readahead_count(rac);
371 do_mpage_readpage(args: &args);
372 /*
373 * If read ahead failed synchronously, it may cause by removed
374 * device, or some filesystem metadata error.
375 */
376 if (!folio_test_locked(folio) && !folio_test_uptodate(folio))
377 break;
378 }
379 if (args.bio)
380 mpage_bio_submit_read(bio: args.bio);
381}
382EXPORT_SYMBOL(mpage_readahead);
383
384/*
385 * This isn't called much at all
386 */
387int mpage_read_folio(struct folio *folio, get_block_t get_block)
388{
389 struct mpage_readpage_args args = {
390 .folio = folio,
391 .nr_pages = folio_nr_pages(folio),
392 .get_block = get_block,
393 };
394
395 do_mpage_readpage(args: &args);
396 if (args.bio)
397 mpage_bio_submit_read(bio: args.bio);
398 return 0;
399}
400EXPORT_SYMBOL(mpage_read_folio);
401
402/*
403 * Writing is not so simple.
404 *
405 * If the page has buffers then they will be used for obtaining the disk
406 * mapping. We only support pages which are fully mapped-and-dirty, with a
407 * special case for pages which are unmapped at the end: end-of-file.
408 *
409 * If the page has no buffers (preferred) then the page is mapped here.
410 *
411 * If all blocks are found to be contiguous then the page can go into the
412 * BIO. Otherwise fall back to the mapping's writepage().
413 *
414 * FIXME: This code wants an estimate of how many pages are still to be
415 * written, so it can intelligently allocate a suitably-sized BIO. For now,
416 * just allocate full-size (16-page) BIOs.
417 */
418
419struct mpage_data {
420 struct bio *bio;
421 sector_t last_block_in_bio;
422 get_block_t *get_block;
423};
424
425/*
426 * We have our BIO, so we can now mark the buffers clean. Make
427 * sure to only clean buffers which we know we'll be writing.
428 */
429static void clean_buffers(struct folio *folio, unsigned first_unmapped)
430{
431 unsigned buffer_counter = 0;
432 struct buffer_head *bh, *head = folio_buffers(folio);
433
434 if (!head)
435 return;
436 bh = head;
437
438 do {
439 if (buffer_counter++ == first_unmapped)
440 break;
441 clear_buffer_dirty(bh);
442 bh = bh->b_this_page;
443 } while (bh != head);
444
445 /*
446 * we cannot drop the bh if the page is not uptodate or a concurrent
447 * read_folio would fail to serialize with the bh and it would read from
448 * disk before we reach the platter.
449 */
450 if (buffer_heads_over_limit && folio_test_uptodate(folio))
451 try_to_free_buffers(folio);
452}
453
454static int mpage_write_folio(struct writeback_control *wbc, struct folio *folio,
455 struct mpage_data *mpd)
456{
457 struct bio *bio = mpd->bio;
458 struct address_space *mapping = folio->mapping;
459 struct inode *inode = mapping->host;
460 const unsigned blkbits = inode->i_blkbits;
461 const unsigned blocks_per_folio = folio_size(folio) >> blkbits;
462 sector_t last_block;
463 sector_t block_in_file;
464 sector_t first_block;
465 unsigned page_block;
466 unsigned first_unmapped = blocks_per_folio;
467 struct block_device *bdev = NULL;
468 int boundary = 0;
469 sector_t boundary_block = 0;
470 struct block_device *boundary_bdev = NULL;
471 size_t length;
472 struct buffer_head map_bh;
473 loff_t i_size = i_size_read(inode);
474 int ret = 0;
475 struct buffer_head *head = folio_buffers(folio);
476
477 if (head) {
478 struct buffer_head *bh = head;
479
480 /* If they're all mapped and dirty, do it */
481 page_block = 0;
482 do {
483 BUG_ON(buffer_locked(bh));
484 if (!buffer_mapped(bh)) {
485 /*
486 * unmapped dirty buffers are created by
487 * block_dirty_folio -> mmapped data
488 */
489 if (buffer_dirty(bh))
490 goto confused;
491 if (first_unmapped == blocks_per_folio)
492 first_unmapped = page_block;
493 continue;
494 }
495
496 if (first_unmapped != blocks_per_folio)
497 goto confused; /* hole -> non-hole */
498
499 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
500 goto confused;
501 if (page_block) {
502 if (bh->b_blocknr != first_block + page_block)
503 goto confused;
504 } else {
505 first_block = bh->b_blocknr;
506 }
507 page_block++;
508 boundary = buffer_boundary(bh);
509 if (boundary) {
510 boundary_block = bh->b_blocknr;
511 boundary_bdev = bh->b_bdev;
512 }
513 bdev = bh->b_bdev;
514 } while ((bh = bh->b_this_page) != head);
515
516 if (first_unmapped)
517 goto page_is_mapped;
518
519 /*
520 * Page has buffers, but they are all unmapped. The page was
521 * created by pagein or read over a hole which was handled by
522 * block_read_full_folio(). If this address_space is also
523 * using mpage_readahead then this can rarely happen.
524 */
525 goto confused;
526 }
527
528 /*
529 * The page has no buffers: map it to disk
530 */
531 BUG_ON(!folio_test_uptodate(folio));
532 block_in_file = folio_pos(folio) >> blkbits;
533 /*
534 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
535 * space.
536 */
537 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
538 goto page_is_mapped;
539 last_block = (i_size - 1) >> blkbits;
540 map_bh.b_folio = folio;
541 for (page_block = 0; page_block < blocks_per_folio; ) {
542
543 map_bh.b_state = 0;
544 map_bh.b_size = 1 << blkbits;
545 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
546 goto confused;
547 if (!buffer_mapped(bh: &map_bh))
548 goto confused;
549 if (buffer_new(bh: &map_bh))
550 clean_bdev_bh_alias(bh: &map_bh);
551 if (buffer_boundary(bh: &map_bh)) {
552 boundary_block = map_bh.b_blocknr;
553 boundary_bdev = map_bh.b_bdev;
554 }
555 if (page_block) {
556 if (map_bh.b_blocknr != first_block + page_block)
557 goto confused;
558 } else {
559 first_block = map_bh.b_blocknr;
560 }
561 page_block++;
562 boundary = buffer_boundary(bh: &map_bh);
563 bdev = map_bh.b_bdev;
564 if (block_in_file == last_block)
565 break;
566 block_in_file++;
567 }
568 BUG_ON(page_block == 0);
569
570 first_unmapped = page_block;
571
572page_is_mapped:
573 /* Don't bother writing beyond EOF, truncate will discard the folio */
574 if (folio_pos(folio) >= i_size)
575 goto confused;
576 length = folio_size(folio);
577 if (folio_pos(folio) + length > i_size) {
578 /*
579 * The page straddles i_size. It must be zeroed out on each
580 * and every writepage invocation because it may be mmapped.
581 * "A file is mapped in multiples of the page size. For a file
582 * that is not a multiple of the page size, the remaining memory
583 * is zeroed when mapped, and writes to that region are not
584 * written out to the file."
585 */
586 length = i_size - folio_pos(folio);
587 folio_zero_segment(folio, start: length, xend: folio_size(folio));
588 }
589
590 /*
591 * This page will go to BIO. Do we need to send this BIO off first?
592 */
593 if (bio && mpd->last_block_in_bio != first_block - 1)
594 bio = mpage_bio_submit_write(bio);
595
596alloc_new:
597 if (bio == NULL) {
598 bio = bio_alloc(bdev, BIO_MAX_VECS,
599 opf: REQ_OP_WRITE | wbc_to_write_flags(wbc),
600 GFP_NOFS);
601 bio->bi_iter.bi_sector = first_block << (blkbits - 9);
602 wbc_init_bio(wbc, bio);
603 bio->bi_write_hint = inode->i_write_hint;
604 }
605
606 /*
607 * Must try to add the page before marking the buffer clean or
608 * the confused fail path above (OOM) will be very confused when
609 * it finds all bh marked clean (i.e. it will not write anything)
610 */
611 wbc_account_cgroup_owner(wbc, folio, bytes: folio_size(folio));
612 length = first_unmapped << blkbits;
613 if (!bio_add_folio(bio, folio, len: length, off: 0)) {
614 bio = mpage_bio_submit_write(bio);
615 goto alloc_new;
616 }
617
618 clean_buffers(folio, first_unmapped);
619
620 BUG_ON(folio_test_writeback(folio));
621 folio_start_writeback(folio);
622 folio_unlock(folio);
623 if (boundary || (first_unmapped != blocks_per_folio)) {
624 bio = mpage_bio_submit_write(bio);
625 if (boundary_block) {
626 write_boundary_block(bdev: boundary_bdev,
627 bblock: boundary_block, blocksize: 1 << blkbits);
628 }
629 } else {
630 mpd->last_block_in_bio = first_block + blocks_per_folio - 1;
631 }
632 goto out;
633
634confused:
635 if (bio)
636 bio = mpage_bio_submit_write(bio);
637
638 /*
639 * The caller has a ref on the inode, so *mapping is stable
640 */
641 ret = block_write_full_folio(folio, wbc, get_block: mpd->get_block);
642 mapping_set_error(mapping, error: ret);
643out:
644 mpd->bio = bio;
645 return ret;
646}
647
648/**
649 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
650 * @mapping: address space structure to write
651 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
652 * @get_block: the filesystem's block mapper function.
653 *
654 * This is a library function, which implements the writepages()
655 * address_space_operation.
656 */
657int
658mpage_writepages(struct address_space *mapping,
659 struct writeback_control *wbc, get_block_t get_block)
660{
661 struct mpage_data mpd = {
662 .get_block = get_block,
663 };
664 struct folio *folio = NULL;
665 struct blk_plug plug;
666 int error;
667
668 blk_start_plug(&plug);
669 while ((folio = writeback_iter(mapping, wbc, folio, error: &error)))
670 error = mpage_write_folio(wbc, folio, mpd: &mpd);
671 if (mpd.bio)
672 mpage_bio_submit_write(bio: mpd.bio);
673 blk_finish_plug(&plug);
674 return error;
675}
676EXPORT_SYMBOL(mpage_writepages);
677