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 "internal.h"
32
33 /*
34 * I/O completion handler for multipage BIOs.
35 *
36 * The mpage code never puts partial pages into a BIO (except for end-of-file).
37 * If a page does not map to a contiguous run of blocks then it simply falls
38 * back to block_read_full_folio().
39 *
40 * Why is this? If a page's completion depends on a number of different BIOs
41 * which can complete in any order (or at the same time) then determining the
42 * status of that page is hard. See end_buffer_async_read() for the details.
43 * There is no point in duplicating all that complexity.
44 */
mpage_read_end_io(struct bio * bio)45 static void mpage_read_end_io(struct bio *bio)
46 {
47 struct folio_iter fi;
48 int err = blk_status_to_errno(bio->bi_status);
49
50 bio_for_each_folio_all(fi, bio)
51 folio_end_read(fi.folio, err == 0);
52
53 bio_put(bio);
54 }
55
mpage_write_end_io(struct bio * bio)56 static void mpage_write_end_io(struct bio *bio)
57 {
58 struct folio_iter fi;
59 int err = blk_status_to_errno(bio->bi_status);
60
61 bio_for_each_folio_all(fi, bio) {
62 if (err)
63 mapping_set_error(fi.folio->mapping, err);
64 folio_end_writeback(fi.folio);
65 }
66
67 bio_put(bio);
68 }
69
mpage_bio_submit_read(struct bio * bio)70 static struct bio *mpage_bio_submit_read(struct bio *bio)
71 {
72 bio->bi_end_io = mpage_read_end_io;
73 guard_bio_eod(bio);
74 submit_bio(bio);
75 return NULL;
76 }
77
mpage_bio_submit_write(struct bio * bio)78 static struct bio *mpage_bio_submit_write(struct bio *bio)
79 {
80 bio->bi_end_io = mpage_write_end_io;
81 guard_bio_eod(bio);
82 submit_bio(bio);
83 return NULL;
84 }
85
86 /*
87 * support function for mpage_readahead. The fs supplied get_block might
88 * return an up to date buffer. This is used to map that buffer into
89 * the page, which allows read_folio to avoid triggering a duplicate call
90 * to get_block.
91 *
92 * The idea is to avoid adding buffers to pages that don't already have
93 * them. So when the buffer is up to date and the page size == block size,
94 * this marks the page up to date instead of adding new buffers.
95 */
map_buffer_to_folio(struct folio * folio,struct buffer_head * bh,int page_block)96 static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
97 int page_block)
98 {
99 struct inode *inode = folio->mapping->host;
100 struct buffer_head *page_bh, *head;
101 int block = 0;
102
103 head = folio_buffers(folio);
104 if (!head) {
105 /*
106 * don't make any buffers if there is only one buffer on
107 * the folio and the folio just needs to be set up to date
108 */
109 if (inode->i_blkbits == folio_shift(folio) &&
110 buffer_uptodate(bh)) {
111 folio_mark_uptodate(folio);
112 return;
113 }
114 head = create_empty_buffers(folio, i_blocksize(inode), 0);
115 }
116
117 page_bh = head;
118 do {
119 if (block == page_block) {
120 page_bh->b_state = bh->b_state;
121 page_bh->b_bdev = bh->b_bdev;
122 page_bh->b_blocknr = bh->b_blocknr;
123 break;
124 }
125 page_bh = page_bh->b_this_page;
126 block++;
127 } while (page_bh != head);
128 }
129
130 struct mpage_readpage_args {
131 struct bio *bio;
132 struct folio *folio;
133 unsigned int nr_pages;
134 bool is_readahead;
135 sector_t last_block_in_bio;
136 struct buffer_head map_bh;
137 unsigned long first_logical_block;
138 get_block_t *get_block;
139 };
140
141 /*
142 * This is the worker routine which does all the work of mapping the disk
143 * blocks and constructs largest possible bios, submits them for IO if the
144 * blocks are not contiguous on the disk.
145 *
146 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
147 * represent the validity of its disk mapping and to decide when to do the next
148 * get_block() call.
149 */
do_mpage_readpage(struct mpage_readpage_args * args)150 static void do_mpage_readpage(struct mpage_readpage_args *args)
151 {
152 struct folio *folio = args->folio;
153 struct inode *inode = folio->mapping->host;
154 const unsigned blkbits = inode->i_blkbits;
155 const unsigned blocks_per_folio = folio_size(folio) >> blkbits;
156 const unsigned blocksize = 1 << blkbits;
157 struct buffer_head *map_bh = &args->map_bh;
158 sector_t block_in_file;
159 sector_t last_block;
160 sector_t last_block_in_file;
161 sector_t first_block;
162 unsigned page_block;
163 unsigned first_hole = blocks_per_folio;
164 struct block_device *bdev = NULL;
165 int length;
166 int fully_mapped = 1;
167 blk_opf_t opf = REQ_OP_READ;
168 unsigned nblocks;
169 unsigned relative_block;
170 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
171
172 if (args->is_readahead) {
173 opf |= REQ_RAHEAD;
174 gfp |= __GFP_NORETRY | __GFP_NOWARN;
175 }
176
177 if (folio_buffers(folio))
178 goto confused;
179
180 block_in_file = folio_pos(folio) >> blkbits;
181 last_block = block_in_file + ((args->nr_pages * PAGE_SIZE) >> blkbits);
182 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
183 if (last_block > last_block_in_file)
184 last_block = last_block_in_file;
185 page_block = 0;
186
187 /*
188 * Map blocks using the result from the previous get_blocks call first.
189 */
190 nblocks = map_bh->b_size >> blkbits;
191 if (buffer_mapped(map_bh) &&
192 block_in_file > args->first_logical_block &&
193 block_in_file < (args->first_logical_block + nblocks)) {
194 unsigned map_offset = block_in_file - args->first_logical_block;
195 unsigned last = nblocks - map_offset;
196
197 first_block = map_bh->b_blocknr + map_offset;
198 for (relative_block = 0; ; relative_block++) {
199 if (relative_block == last) {
200 clear_buffer_mapped(map_bh);
201 break;
202 }
203 if (page_block == blocks_per_folio)
204 break;
205 page_block++;
206 block_in_file++;
207 }
208 bdev = map_bh->b_bdev;
209 }
210
211 /*
212 * Then do more get_blocks calls until we are done with this folio.
213 */
214 map_bh->b_folio = folio;
215 while (page_block < blocks_per_folio) {
216 map_bh->b_state = 0;
217 map_bh->b_size = 0;
218
219 if (block_in_file < last_block) {
220 map_bh->b_size = (last_block-block_in_file) << blkbits;
221 if (args->get_block(inode, block_in_file, map_bh, 0))
222 goto confused;
223 args->first_logical_block = block_in_file;
224 }
225
226 if (!buffer_mapped(map_bh)) {
227 fully_mapped = 0;
228 if (first_hole == blocks_per_folio)
229 first_hole = page_block;
230 page_block++;
231 block_in_file++;
232 continue;
233 }
234
235 /* some filesystems will copy data into the page during
236 * the get_block call, in which case we don't want to
237 * read it again. map_buffer_to_folio copies the data
238 * we just collected from get_block into the folio's buffers
239 * so read_folio doesn't have to repeat the get_block call
240 */
241 if (buffer_uptodate(map_bh)) {
242 map_buffer_to_folio(folio, map_bh, page_block);
243 goto confused;
244 }
245
246 if (first_hole != blocks_per_folio)
247 goto confused; /* hole -> non-hole */
248
249 /* Contiguous blocks? */
250 if (!page_block)
251 first_block = map_bh->b_blocknr;
252 else if (first_block + page_block != map_bh->b_blocknr)
253 goto confused;
254 nblocks = map_bh->b_size >> blkbits;
255 for (relative_block = 0; ; relative_block++) {
256 if (relative_block == nblocks) {
257 clear_buffer_mapped(map_bh);
258 break;
259 } else if (page_block == blocks_per_folio)
260 break;
261 page_block++;
262 block_in_file++;
263 }
264 bdev = map_bh->b_bdev;
265 }
266
267 if (first_hole != blocks_per_folio) {
268 folio_zero_segment(folio, first_hole << blkbits, folio_size(folio));
269 if (first_hole == 0) {
270 folio_mark_uptodate(folio);
271 folio_unlock(folio);
272 goto out;
273 }
274 } else if (fully_mapped) {
275 folio_set_mappedtodisk(folio);
276 }
277
278 /*
279 * This folio will go to BIO. Do we need to send this BIO off first?
280 */
281 if (args->bio && (args->last_block_in_bio != first_block - 1))
282 args->bio = mpage_bio_submit_read(args->bio);
283
284 alloc_new:
285 if (args->bio == NULL) {
286 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
287 gfp);
288 if (args->bio == NULL)
289 goto confused;
290 args->bio->bi_iter.bi_sector = first_block << (blkbits - 9);
291 }
292
293 length = first_hole << blkbits;
294 if (!bio_add_folio(args->bio, folio, length, 0)) {
295 args->bio = mpage_bio_submit_read(args->bio);
296 goto alloc_new;
297 }
298
299 relative_block = block_in_file - args->first_logical_block;
300 nblocks = map_bh->b_size >> blkbits;
301 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
302 (first_hole != blocks_per_folio))
303 args->bio = mpage_bio_submit_read(args->bio);
304 else
305 args->last_block_in_bio = first_block + blocks_per_folio - 1;
306 out:
307 return;
308
309 confused:
310 if (args->bio)
311 args->bio = mpage_bio_submit_read(args->bio);
312 if (!folio_test_uptodate(folio))
313 block_read_full_folio(folio, args->get_block);
314 else
315 folio_unlock(folio);
316 goto out;
317 }
318
319 /**
320 * mpage_readahead - start reads against pages
321 * @rac: Describes which pages to read.
322 * @get_block: The filesystem's block mapper function.
323 *
324 * This function walks the pages and the blocks within each page, building and
325 * emitting large BIOs.
326 *
327 * If anything unusual happens, such as:
328 *
329 * - encountering a page which has buffers
330 * - encountering a page which has a non-hole after a hole
331 * - encountering a page with non-contiguous blocks
332 *
333 * then this code just gives up and calls the buffer_head-based read function.
334 * It does handle a page which has holes at the end - that is a common case:
335 * the end-of-file on blocksize < PAGE_SIZE setups.
336 *
337 * BH_Boundary explanation:
338 *
339 * There is a problem. The mpage read code assembles several pages, gets all
340 * their disk mappings, and then submits them all. That's fine, but obtaining
341 * the disk mappings may require I/O. Reads of indirect blocks, for example.
342 *
343 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
344 * submitted in the following order:
345 *
346 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
347 *
348 * because the indirect block has to be read to get the mappings of blocks
349 * 13,14,15,16. Obviously, this impacts performance.
350 *
351 * So what we do it to allow the filesystem's get_block() function to set
352 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
353 * after this one will require I/O against a block which is probably close to
354 * this one. So you should push what I/O you have currently accumulated.
355 *
356 * This all causes the disk requests to be issued in the correct order.
357 */
mpage_readahead(struct readahead_control * rac,get_block_t get_block)358 void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
359 {
360 struct folio *folio;
361 struct mpage_readpage_args args = {
362 .get_block = get_block,
363 .is_readahead = true,
364 };
365
366 while ((folio = readahead_folio(rac))) {
367 prefetchw(&folio->flags);
368 args.folio = folio;
369 args.nr_pages = readahead_count(rac);
370 do_mpage_readpage(&args);
371 /*
372 * If read ahead failed synchronously, it may cause by removed
373 * device, or some filesystem metadata error.
374 */
375 if (!folio_test_locked(folio) && !folio_test_uptodate(folio))
376 break;
377 }
378 if (args.bio)
379 mpage_bio_submit_read(args.bio);
380 }
381 EXPORT_SYMBOL(mpage_readahead);
382
383 /*
384 * This isn't called much at all
385 */
mpage_read_folio(struct folio * folio,get_block_t get_block)386 int mpage_read_folio(struct folio *folio, get_block_t get_block)
387 {
388 struct mpage_readpage_args args = {
389 .folio = folio,
390 .nr_pages = folio_nr_pages(folio),
391 .get_block = get_block,
392 };
393
394 do_mpage_readpage(&args);
395 if (args.bio)
396 mpage_bio_submit_read(args.bio);
397 return 0;
398 }
399 EXPORT_SYMBOL(mpage_read_folio);
400
401 /*
402 * Writing is not so simple.
403 *
404 * If the page has buffers then they will be used for obtaining the disk
405 * mapping. We only support pages which are fully mapped-and-dirty, with a
406 * special case for pages which are unmapped at the end: end-of-file.
407 *
408 * If the page has no buffers (preferred) then the page is mapped here.
409 *
410 * If all blocks are found to be contiguous then the page can go into the
411 * BIO. Otherwise fall back to the mapping's writepage().
412 *
413 * FIXME: This code wants an estimate of how many pages are still to be
414 * written, so it can intelligently allocate a suitably-sized BIO. For now,
415 * just allocate full-size (16-page) BIOs.
416 */
417
418 struct mpage_data {
419 struct bio *bio;
420 sector_t last_block_in_bio;
421 get_block_t *get_block;
422 };
423
424 /*
425 * We have our BIO, so we can now mark the buffers clean. Make
426 * sure to only clean buffers which we know we'll be writing.
427 */
clean_buffers(struct folio * folio,unsigned first_unmapped)428 static void clean_buffers(struct folio *folio, unsigned first_unmapped)
429 {
430 unsigned buffer_counter = 0;
431 struct buffer_head *bh, *head = folio_buffers(folio);
432
433 if (!head)
434 return;
435 bh = head;
436
437 do {
438 if (buffer_counter++ == first_unmapped)
439 break;
440 clear_buffer_dirty(bh);
441 bh = bh->b_this_page;
442 } while (bh != head);
443
444 /*
445 * we cannot drop the bh if the page is not uptodate or a concurrent
446 * read_folio would fail to serialize with the bh and it would read from
447 * disk before we reach the platter.
448 */
449 if (buffer_heads_over_limit && folio_test_uptodate(folio))
450 try_to_free_buffers(folio);
451 }
452
mpage_write_folio(struct writeback_control * wbc,struct folio * folio,struct mpage_data * mpd)453 static int mpage_write_folio(struct writeback_control *wbc, struct folio *folio,
454 struct mpage_data *mpd)
455 {
456 struct bio *bio = mpd->bio;
457 struct address_space *mapping = folio->mapping;
458 struct inode *inode = mapping->host;
459 const unsigned blkbits = inode->i_blkbits;
460 const unsigned blocks_per_folio = folio_size(folio) >> blkbits;
461 sector_t last_block;
462 sector_t block_in_file;
463 sector_t first_block;
464 unsigned page_block;
465 unsigned first_unmapped = blocks_per_folio;
466 struct block_device *bdev = NULL;
467 int boundary = 0;
468 sector_t boundary_block = 0;
469 struct block_device *boundary_bdev = NULL;
470 size_t length;
471 struct buffer_head map_bh;
472 loff_t i_size = i_size_read(inode);
473 int ret = 0;
474 struct buffer_head *head = folio_buffers(folio);
475
476 if (head) {
477 struct buffer_head *bh = head;
478
479 /* If they're all mapped and dirty, do it */
480 page_block = 0;
481 do {
482 BUG_ON(buffer_locked(bh));
483 if (!buffer_mapped(bh)) {
484 /*
485 * unmapped dirty buffers are created by
486 * block_dirty_folio -> mmapped data
487 */
488 if (buffer_dirty(bh))
489 goto confused;
490 if (first_unmapped == blocks_per_folio)
491 first_unmapped = page_block;
492 continue;
493 }
494
495 if (first_unmapped != blocks_per_folio)
496 goto confused; /* hole -> non-hole */
497
498 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
499 goto confused;
500 if (page_block) {
501 if (bh->b_blocknr != first_block + page_block)
502 goto confused;
503 } else {
504 first_block = bh->b_blocknr;
505 }
506 page_block++;
507 boundary = buffer_boundary(bh);
508 if (boundary) {
509 boundary_block = bh->b_blocknr;
510 boundary_bdev = bh->b_bdev;
511 }
512 bdev = bh->b_bdev;
513 } while ((bh = bh->b_this_page) != head);
514
515 if (first_unmapped)
516 goto page_is_mapped;
517
518 /*
519 * Page has buffers, but they are all unmapped. The page was
520 * created by pagein or read over a hole which was handled by
521 * block_read_full_folio(). If this address_space is also
522 * using mpage_readahead then this can rarely happen.
523 */
524 goto confused;
525 }
526
527 /*
528 * The page has no buffers: map it to disk
529 */
530 BUG_ON(!folio_test_uptodate(folio));
531 block_in_file = folio_pos(folio) >> blkbits;
532 /*
533 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
534 * space.
535 */
536 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
537 goto page_is_mapped;
538 last_block = (i_size - 1) >> blkbits;
539 map_bh.b_folio = folio;
540 for (page_block = 0; page_block < blocks_per_folio; ) {
541
542 map_bh.b_state = 0;
543 map_bh.b_size = 1 << blkbits;
544 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
545 goto confused;
546 if (!buffer_mapped(&map_bh))
547 goto confused;
548 if (buffer_new(&map_bh))
549 clean_bdev_bh_alias(&map_bh);
550 if (buffer_boundary(&map_bh)) {
551 boundary_block = map_bh.b_blocknr;
552 boundary_bdev = map_bh.b_bdev;
553 }
554 if (page_block) {
555 if (map_bh.b_blocknr != first_block + page_block)
556 goto confused;
557 } else {
558 first_block = map_bh.b_blocknr;
559 }
560 page_block++;
561 boundary = buffer_boundary(&map_bh);
562 bdev = map_bh.b_bdev;
563 if (block_in_file == last_block)
564 break;
565 block_in_file++;
566 }
567 BUG_ON(page_block == 0);
568
569 first_unmapped = page_block;
570
571 page_is_mapped:
572 /* Don't bother writing beyond EOF, truncate will discard the folio */
573 if (folio_pos(folio) >= i_size)
574 goto confused;
575 length = folio_size(folio);
576 if (folio_pos(folio) + length > i_size) {
577 /*
578 * The page straddles i_size. It must be zeroed out on each
579 * and every writepage invocation because it may be mmapped.
580 * "A file is mapped in multiples of the page size. For a file
581 * that is not a multiple of the page size, the remaining memory
582 * is zeroed when mapped, and writes to that region are not
583 * written out to the file."
584 */
585 length = i_size - folio_pos(folio);
586 folio_zero_segment(folio, length, folio_size(folio));
587 }
588
589 /*
590 * This page will go to BIO. Do we need to send this BIO off first?
591 */
592 if (bio && mpd->last_block_in_bio != first_block - 1)
593 bio = mpage_bio_submit_write(bio);
594
595 alloc_new:
596 if (bio == NULL) {
597 bio = bio_alloc(bdev, BIO_MAX_VECS,
598 REQ_OP_WRITE | wbc_to_write_flags(wbc),
599 GFP_NOFS);
600 bio->bi_iter.bi_sector = first_block << (blkbits - 9);
601 wbc_init_bio(wbc, bio);
602 bio->bi_write_hint = inode->i_write_hint;
603 }
604
605 /*
606 * Must try to add the page before marking the buffer clean or
607 * the confused fail path above (OOM) will be very confused when
608 * it finds all bh marked clean (i.e. it will not write anything)
609 */
610 wbc_account_cgroup_owner(wbc, folio, folio_size(folio));
611 length = first_unmapped << blkbits;
612 if (!bio_add_folio(bio, folio, length, 0)) {
613 bio = mpage_bio_submit_write(bio);
614 goto alloc_new;
615 }
616
617 clean_buffers(folio, first_unmapped);
618
619 BUG_ON(folio_test_writeback(folio));
620 folio_start_writeback(folio);
621 folio_unlock(folio);
622 if (boundary || (first_unmapped != blocks_per_folio)) {
623 bio = mpage_bio_submit_write(bio);
624 if (boundary_block) {
625 write_boundary_block(boundary_bdev,
626 boundary_block, 1 << blkbits);
627 }
628 } else {
629 mpd->last_block_in_bio = first_block + blocks_per_folio - 1;
630 }
631 goto out;
632
633 confused:
634 if (bio)
635 bio = mpage_bio_submit_write(bio);
636
637 /*
638 * The caller has a ref on the inode, so *mapping is stable
639 */
640 ret = block_write_full_folio(folio, wbc, mpd->get_block);
641 mapping_set_error(mapping, ret);
642 out:
643 mpd->bio = bio;
644 return ret;
645 }
646
647 /**
648 * __mpage_writepages - walk the list of dirty pages of the given address space
649 * & 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 * @write_folio: handler to call for each folio before calling
654 * mpage_write_folio()
655 *
656 * This is a library function, which implements the writepages()
657 * address_space_operation. It calls @write_folio handler for each folio. If
658 * the handler returns value > 0, it calls mpage_write_folio() to do the
659 * folio writeback.
660 */
661 int
__mpage_writepages(struct address_space * mapping,struct writeback_control * wbc,get_block_t get_block,int (* write_folio)(struct folio * folio,struct writeback_control * wbc))662 __mpage_writepages(struct address_space *mapping,
663 struct writeback_control *wbc, get_block_t get_block,
664 int (*write_folio)(struct folio *folio,
665 struct writeback_control *wbc))
666 {
667 struct mpage_data mpd = {
668 .get_block = get_block,
669 };
670 struct folio *folio = NULL;
671 struct blk_plug plug;
672 int error;
673
674 blk_start_plug(&plug);
675 while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
676 if (write_folio) {
677 error = write_folio(folio, wbc);
678 /*
679 * == 0 means folio is handled, < 0 means error. In
680 * both cases hand back control to writeback_iter()
681 */
682 if (error <= 0)
683 continue;
684 /* Let mpage_write_folio() handle the folio. */
685 }
686 error = mpage_write_folio(wbc, folio, &mpd);
687 }
688 if (mpd.bio)
689 mpage_bio_submit_write(mpd.bio);
690 blk_finish_plug(&plug);
691 return error;
692 }
693 EXPORT_SYMBOL(__mpage_writepages);
694