Loading...
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 <linux/cleancache.h>
33#include "internal.h"
34
35/*
36 * I/O completion handler for multipage BIOs.
37 *
38 * The mpage code never puts partial pages into a BIO (except for end-of-file).
39 * If a page does not map to a contiguous run of blocks then it simply falls
40 * back to block_read_full_page().
41 *
42 * Why is this? If a page's completion depends on a number of different BIOs
43 * which can complete in any order (or at the same time) then determining the
44 * status of that page is hard. See end_buffer_async_read() for the details.
45 * There is no point in duplicating all that complexity.
46 */
47static void mpage_end_io(struct bio *bio)
48{
49 struct bio_vec *bv;
50 struct bvec_iter_all iter_all;
51
52 bio_for_each_segment_all(bv, bio, iter_all) {
53 struct page *page = bv->bv_page;
54 page_endio(page, bio_op(bio),
55 blk_status_to_errno(bio->bi_status));
56 }
57
58 bio_put(bio);
59}
60
61static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
62{
63 bio->bi_end_io = mpage_end_io;
64 bio_set_op_attrs(bio, op, op_flags);
65 guard_bio_eod(op, bio);
66 submit_bio(bio);
67 return NULL;
68}
69
70static struct bio *
71mpage_alloc(struct block_device *bdev,
72 sector_t first_sector, int nr_vecs,
73 gfp_t gfp_flags)
74{
75 struct bio *bio;
76
77 /* Restrict the given (page cache) mask for slab allocations */
78 gfp_flags &= GFP_KERNEL;
79 bio = bio_alloc(gfp_flags, nr_vecs);
80
81 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
82 while (!bio && (nr_vecs /= 2))
83 bio = bio_alloc(gfp_flags, nr_vecs);
84 }
85
86 if (bio) {
87 bio_set_dev(bio, bdev);
88 bio->bi_iter.bi_sector = first_sector;
89 }
90 return bio;
91}
92
93/*
94 * support function for mpage_readpages. The fs supplied get_block might
95 * return an up to date buffer. This is used to map that buffer into
96 * the page, which allows readpage to avoid triggering a duplicate call
97 * to get_block.
98 *
99 * The idea is to avoid adding buffers to pages that don't already have
100 * them. So when the buffer is up to date and the page size == block size,
101 * this marks the page up to date instead of adding new buffers.
102 */
103static void
104map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
105{
106 struct inode *inode = page->mapping->host;
107 struct buffer_head *page_bh, *head;
108 int block = 0;
109
110 if (!page_has_buffers(page)) {
111 /*
112 * don't make any buffers if there is only one buffer on
113 * the page and the page just needs to be set up to date
114 */
115 if (inode->i_blkbits == PAGE_SHIFT &&
116 buffer_uptodate(bh)) {
117 SetPageUptodate(page);
118 return;
119 }
120 create_empty_buffers(page, i_blocksize(inode), 0);
121 }
122 head = page_buffers(page);
123 page_bh = head;
124 do {
125 if (block == page_block) {
126 page_bh->b_state = bh->b_state;
127 page_bh->b_bdev = bh->b_bdev;
128 page_bh->b_blocknr = bh->b_blocknr;
129 break;
130 }
131 page_bh = page_bh->b_this_page;
132 block++;
133 } while (page_bh != head);
134}
135
136struct mpage_readpage_args {
137 struct bio *bio;
138 struct page *page;
139 unsigned int nr_pages;
140 bool is_readahead;
141 sector_t last_block_in_bio;
142 struct buffer_head map_bh;
143 unsigned long first_logical_block;
144 get_block_t *get_block;
145};
146
147/*
148 * This is the worker routine which does all the work of mapping the disk
149 * blocks and constructs largest possible bios, submits them for IO if the
150 * blocks are not contiguous on the disk.
151 *
152 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
153 * represent the validity of its disk mapping and to decide when to do the next
154 * get_block() call.
155 */
156static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
157{
158 struct page *page = args->page;
159 struct inode *inode = page->mapping->host;
160 const unsigned blkbits = inode->i_blkbits;
161 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
162 const unsigned blocksize = 1 << blkbits;
163 struct buffer_head *map_bh = &args->map_bh;
164 sector_t block_in_file;
165 sector_t last_block;
166 sector_t last_block_in_file;
167 sector_t blocks[MAX_BUF_PER_PAGE];
168 unsigned page_block;
169 unsigned first_hole = blocks_per_page;
170 struct block_device *bdev = NULL;
171 int length;
172 int fully_mapped = 1;
173 int op_flags;
174 unsigned nblocks;
175 unsigned relative_block;
176 gfp_t gfp;
177
178 if (args->is_readahead) {
179 op_flags = REQ_RAHEAD;
180 gfp = readahead_gfp_mask(page->mapping);
181 } else {
182 op_flags = 0;
183 gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
184 }
185
186 if (page_has_buffers(page))
187 goto confused;
188
189 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
190 last_block = block_in_file + args->nr_pages * blocks_per_page;
191 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
192 if (last_block > last_block_in_file)
193 last_block = last_block_in_file;
194 page_block = 0;
195
196 /*
197 * Map blocks using the result from the previous get_blocks call first.
198 */
199 nblocks = map_bh->b_size >> blkbits;
200 if (buffer_mapped(map_bh) &&
201 block_in_file > args->first_logical_block &&
202 block_in_file < (args->first_logical_block + nblocks)) {
203 unsigned map_offset = block_in_file - args->first_logical_block;
204 unsigned last = nblocks - map_offset;
205
206 for (relative_block = 0; ; relative_block++) {
207 if (relative_block == last) {
208 clear_buffer_mapped(map_bh);
209 break;
210 }
211 if (page_block == blocks_per_page)
212 break;
213 blocks[page_block] = map_bh->b_blocknr + map_offset +
214 relative_block;
215 page_block++;
216 block_in_file++;
217 }
218 bdev = map_bh->b_bdev;
219 }
220
221 /*
222 * Then do more get_blocks calls until we are done with this page.
223 */
224 map_bh->b_page = page;
225 while (page_block < blocks_per_page) {
226 map_bh->b_state = 0;
227 map_bh->b_size = 0;
228
229 if (block_in_file < last_block) {
230 map_bh->b_size = (last_block-block_in_file) << blkbits;
231 if (args->get_block(inode, block_in_file, map_bh, 0))
232 goto confused;
233 args->first_logical_block = block_in_file;
234 }
235
236 if (!buffer_mapped(map_bh)) {
237 fully_mapped = 0;
238 if (first_hole == blocks_per_page)
239 first_hole = page_block;
240 page_block++;
241 block_in_file++;
242 continue;
243 }
244
245 /* some filesystems will copy data into the page during
246 * the get_block call, in which case we don't want to
247 * read it again. map_buffer_to_page copies the data
248 * we just collected from get_block into the page's buffers
249 * so readpage doesn't have to repeat the get_block call
250 */
251 if (buffer_uptodate(map_bh)) {
252 map_buffer_to_page(page, map_bh, page_block);
253 goto confused;
254 }
255
256 if (first_hole != blocks_per_page)
257 goto confused; /* hole -> non-hole */
258
259 /* Contiguous blocks? */
260 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
261 goto confused;
262 nblocks = map_bh->b_size >> blkbits;
263 for (relative_block = 0; ; relative_block++) {
264 if (relative_block == nblocks) {
265 clear_buffer_mapped(map_bh);
266 break;
267 } else if (page_block == blocks_per_page)
268 break;
269 blocks[page_block] = map_bh->b_blocknr+relative_block;
270 page_block++;
271 block_in_file++;
272 }
273 bdev = map_bh->b_bdev;
274 }
275
276 if (first_hole != blocks_per_page) {
277 zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
278 if (first_hole == 0) {
279 SetPageUptodate(page);
280 unlock_page(page);
281 goto out;
282 }
283 } else if (fully_mapped) {
284 SetPageMappedToDisk(page);
285 }
286
287 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
288 cleancache_get_page(page) == 0) {
289 SetPageUptodate(page);
290 goto confused;
291 }
292
293 /*
294 * This page will go to BIO. Do we need to send this BIO off first?
295 */
296 if (args->bio && (args->last_block_in_bio != blocks[0] - 1))
297 args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
298
299alloc_new:
300 if (args->bio == NULL) {
301 if (first_hole == blocks_per_page) {
302 if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
303 page))
304 goto out;
305 }
306 args->bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
307 min_t(int, args->nr_pages,
308 BIO_MAX_PAGES),
309 gfp);
310 if (args->bio == NULL)
311 goto confused;
312 }
313
314 length = first_hole << blkbits;
315 if (bio_add_page(args->bio, page, length, 0) < length) {
316 args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
317 goto alloc_new;
318 }
319
320 relative_block = block_in_file - args->first_logical_block;
321 nblocks = map_bh->b_size >> blkbits;
322 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
323 (first_hole != blocks_per_page))
324 args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
325 else
326 args->last_block_in_bio = blocks[blocks_per_page - 1];
327out:
328 return args->bio;
329
330confused:
331 if (args->bio)
332 args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
333 if (!PageUptodate(page))
334 block_read_full_page(page, args->get_block);
335 else
336 unlock_page(page);
337 goto out;
338}
339
340/**
341 * mpage_readpages - populate an address space with some pages & start reads against them
342 * @mapping: the address_space
343 * @pages: The address of a list_head which contains the target pages. These
344 * pages have their ->index populated and are otherwise uninitialised.
345 * The page at @pages->prev has the lowest file offset, and reads should be
346 * issued in @pages->prev to @pages->next order.
347 * @nr_pages: The number of pages at *@pages
348 * @get_block: The filesystem's block mapper function.
349 *
350 * This function walks the pages and the blocks within each page, building and
351 * emitting large BIOs.
352 *
353 * If anything unusual happens, such as:
354 *
355 * - encountering a page which has buffers
356 * - encountering a page which has a non-hole after a hole
357 * - encountering a page with non-contiguous blocks
358 *
359 * then this code just gives up and calls the buffer_head-based read function.
360 * It does handle a page which has holes at the end - that is a common case:
361 * the end-of-file on blocksize < PAGE_SIZE setups.
362 *
363 * BH_Boundary explanation:
364 *
365 * There is a problem. The mpage read code assembles several pages, gets all
366 * their disk mappings, and then submits them all. That's fine, but obtaining
367 * the disk mappings may require I/O. Reads of indirect blocks, for example.
368 *
369 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
370 * submitted in the following order:
371 *
372 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
373 *
374 * because the indirect block has to be read to get the mappings of blocks
375 * 13,14,15,16. Obviously, this impacts performance.
376 *
377 * So what we do it to allow the filesystem's get_block() function to set
378 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
379 * after this one will require I/O against a block which is probably close to
380 * this one. So you should push what I/O you have currently accumulated.
381 *
382 * This all causes the disk requests to be issued in the correct order.
383 */
384int
385mpage_readpages(struct address_space *mapping, struct list_head *pages,
386 unsigned nr_pages, get_block_t get_block)
387{
388 struct mpage_readpage_args args = {
389 .get_block = get_block,
390 .is_readahead = true,
391 };
392 unsigned page_idx;
393
394 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
395 struct page *page = lru_to_page(pages);
396
397 prefetchw(&page->flags);
398 list_del(&page->lru);
399 if (!add_to_page_cache_lru(page, mapping,
400 page->index,
401 readahead_gfp_mask(mapping))) {
402 args.page = page;
403 args.nr_pages = nr_pages - page_idx;
404 args.bio = do_mpage_readpage(&args);
405 }
406 put_page(page);
407 }
408 BUG_ON(!list_empty(pages));
409 if (args.bio)
410 mpage_bio_submit(REQ_OP_READ, REQ_RAHEAD, args.bio);
411 return 0;
412}
413EXPORT_SYMBOL(mpage_readpages);
414
415/*
416 * This isn't called much at all
417 */
418int mpage_readpage(struct page *page, get_block_t get_block)
419{
420 struct mpage_readpage_args args = {
421 .page = page,
422 .nr_pages = 1,
423 .get_block = get_block,
424 };
425
426 args.bio = do_mpage_readpage(&args);
427 if (args.bio)
428 mpage_bio_submit(REQ_OP_READ, 0, args.bio);
429 return 0;
430}
431EXPORT_SYMBOL(mpage_readpage);
432
433/*
434 * Writing is not so simple.
435 *
436 * If the page has buffers then they will be used for obtaining the disk
437 * mapping. We only support pages which are fully mapped-and-dirty, with a
438 * special case for pages which are unmapped at the end: end-of-file.
439 *
440 * If the page has no buffers (preferred) then the page is mapped here.
441 *
442 * If all blocks are found to be contiguous then the page can go into the
443 * BIO. Otherwise fall back to the mapping's writepage().
444 *
445 * FIXME: This code wants an estimate of how many pages are still to be
446 * written, so it can intelligently allocate a suitably-sized BIO. For now,
447 * just allocate full-size (16-page) BIOs.
448 */
449
450struct mpage_data {
451 struct bio *bio;
452 sector_t last_block_in_bio;
453 get_block_t *get_block;
454 unsigned use_writepage;
455};
456
457/*
458 * We have our BIO, so we can now mark the buffers clean. Make
459 * sure to only clean buffers which we know we'll be writing.
460 */
461static void clean_buffers(struct page *page, unsigned first_unmapped)
462{
463 unsigned buffer_counter = 0;
464 struct buffer_head *bh, *head;
465 if (!page_has_buffers(page))
466 return;
467 head = page_buffers(page);
468 bh = head;
469
470 do {
471 if (buffer_counter++ == first_unmapped)
472 break;
473 clear_buffer_dirty(bh);
474 bh = bh->b_this_page;
475 } while (bh != head);
476
477 /*
478 * we cannot drop the bh if the page is not uptodate or a concurrent
479 * readpage would fail to serialize with the bh and it would read from
480 * disk before we reach the platter.
481 */
482 if (buffer_heads_over_limit && PageUptodate(page))
483 try_to_free_buffers(page);
484}
485
486/*
487 * For situations where we want to clean all buffers attached to a page.
488 * We don't need to calculate how many buffers are attached to the page,
489 * we just need to specify a number larger than the maximum number of buffers.
490 */
491void clean_page_buffers(struct page *page)
492{
493 clean_buffers(page, ~0U);
494}
495
496static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
497 void *data)
498{
499 struct mpage_data *mpd = data;
500 struct bio *bio = mpd->bio;
501 struct address_space *mapping = page->mapping;
502 struct inode *inode = page->mapping->host;
503 const unsigned blkbits = inode->i_blkbits;
504 unsigned long end_index;
505 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
506 sector_t last_block;
507 sector_t block_in_file;
508 sector_t blocks[MAX_BUF_PER_PAGE];
509 unsigned page_block;
510 unsigned first_unmapped = blocks_per_page;
511 struct block_device *bdev = NULL;
512 int boundary = 0;
513 sector_t boundary_block = 0;
514 struct block_device *boundary_bdev = NULL;
515 int length;
516 struct buffer_head map_bh;
517 loff_t i_size = i_size_read(inode);
518 int ret = 0;
519 int op_flags = wbc_to_write_flags(wbc);
520
521 if (page_has_buffers(page)) {
522 struct buffer_head *head = page_buffers(page);
523 struct buffer_head *bh = head;
524
525 /* If they're all mapped and dirty, do it */
526 page_block = 0;
527 do {
528 BUG_ON(buffer_locked(bh));
529 if (!buffer_mapped(bh)) {
530 /*
531 * unmapped dirty buffers are created by
532 * __set_page_dirty_buffers -> mmapped data
533 */
534 if (buffer_dirty(bh))
535 goto confused;
536 if (first_unmapped == blocks_per_page)
537 first_unmapped = page_block;
538 continue;
539 }
540
541 if (first_unmapped != blocks_per_page)
542 goto confused; /* hole -> non-hole */
543
544 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
545 goto confused;
546 if (page_block) {
547 if (bh->b_blocknr != blocks[page_block-1] + 1)
548 goto confused;
549 }
550 blocks[page_block++] = bh->b_blocknr;
551 boundary = buffer_boundary(bh);
552 if (boundary) {
553 boundary_block = bh->b_blocknr;
554 boundary_bdev = bh->b_bdev;
555 }
556 bdev = bh->b_bdev;
557 } while ((bh = bh->b_this_page) != head);
558
559 if (first_unmapped)
560 goto page_is_mapped;
561
562 /*
563 * Page has buffers, but they are all unmapped. The page was
564 * created by pagein or read over a hole which was handled by
565 * block_read_full_page(). If this address_space is also
566 * using mpage_readpages then this can rarely happen.
567 */
568 goto confused;
569 }
570
571 /*
572 * The page has no buffers: map it to disk
573 */
574 BUG_ON(!PageUptodate(page));
575 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
576 last_block = (i_size - 1) >> blkbits;
577 map_bh.b_page = page;
578 for (page_block = 0; page_block < blocks_per_page; ) {
579
580 map_bh.b_state = 0;
581 map_bh.b_size = 1 << blkbits;
582 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
583 goto confused;
584 if (buffer_new(&map_bh))
585 clean_bdev_bh_alias(&map_bh);
586 if (buffer_boundary(&map_bh)) {
587 boundary_block = map_bh.b_blocknr;
588 boundary_bdev = map_bh.b_bdev;
589 }
590 if (page_block) {
591 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
592 goto confused;
593 }
594 blocks[page_block++] = map_bh.b_blocknr;
595 boundary = buffer_boundary(&map_bh);
596 bdev = map_bh.b_bdev;
597 if (block_in_file == last_block)
598 break;
599 block_in_file++;
600 }
601 BUG_ON(page_block == 0);
602
603 first_unmapped = page_block;
604
605page_is_mapped:
606 end_index = i_size >> PAGE_SHIFT;
607 if (page->index >= end_index) {
608 /*
609 * The page straddles i_size. It must be zeroed out on each
610 * and every writepage invocation because it may be mmapped.
611 * "A file is mapped in multiples of the page size. For a file
612 * that is not a multiple of the page size, the remaining memory
613 * is zeroed when mapped, and writes to that region are not
614 * written out to the file."
615 */
616 unsigned offset = i_size & (PAGE_SIZE - 1);
617
618 if (page->index > end_index || !offset)
619 goto confused;
620 zero_user_segment(page, offset, PAGE_SIZE);
621 }
622
623 /*
624 * This page will go to BIO. Do we need to send this BIO off first?
625 */
626 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
627 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
628
629alloc_new:
630 if (bio == NULL) {
631 if (first_unmapped == blocks_per_page) {
632 if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
633 page, wbc))
634 goto out;
635 }
636 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
637 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
638 if (bio == NULL)
639 goto confused;
640
641 wbc_init_bio(wbc, bio);
642 bio->bi_write_hint = inode->i_write_hint;
643 }
644
645 /*
646 * Must try to add the page before marking the buffer clean or
647 * the confused fail path above (OOM) will be very confused when
648 * it finds all bh marked clean (i.e. it will not write anything)
649 */
650 wbc_account_cgroup_owner(wbc, page, PAGE_SIZE);
651 length = first_unmapped << blkbits;
652 if (bio_add_page(bio, page, length, 0) < length) {
653 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
654 goto alloc_new;
655 }
656
657 clean_buffers(page, first_unmapped);
658
659 BUG_ON(PageWriteback(page));
660 set_page_writeback(page);
661 unlock_page(page);
662 if (boundary || (first_unmapped != blocks_per_page)) {
663 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
664 if (boundary_block) {
665 write_boundary_block(boundary_bdev,
666 boundary_block, 1 << blkbits);
667 }
668 } else {
669 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
670 }
671 goto out;
672
673confused:
674 if (bio)
675 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
676
677 if (mpd->use_writepage) {
678 ret = mapping->a_ops->writepage(page, wbc);
679 } else {
680 ret = -EAGAIN;
681 goto out;
682 }
683 /*
684 * The caller has a ref on the inode, so *mapping is stable
685 */
686 mapping_set_error(mapping, ret);
687out:
688 mpd->bio = bio;
689 return ret;
690}
691
692/**
693 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
694 * @mapping: address space structure to write
695 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
696 * @get_block: the filesystem's block mapper function.
697 * If this is NULL then use a_ops->writepage. Otherwise, go
698 * direct-to-BIO.
699 *
700 * This is a library function, which implements the writepages()
701 * address_space_operation.
702 *
703 * If a page is already under I/O, generic_writepages() skips it, even
704 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
705 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
706 * and msync() need to guarantee that all the data which was dirty at the time
707 * the call was made get new I/O started against them. If wbc->sync_mode is
708 * WB_SYNC_ALL then we were called for data integrity and we must wait for
709 * existing IO to complete.
710 */
711int
712mpage_writepages(struct address_space *mapping,
713 struct writeback_control *wbc, get_block_t get_block)
714{
715 struct blk_plug plug;
716 int ret;
717
718 blk_start_plug(&plug);
719
720 if (!get_block)
721 ret = generic_writepages(mapping, wbc);
722 else {
723 struct mpage_data mpd = {
724 .bio = NULL,
725 .last_block_in_bio = 0,
726 .get_block = get_block,
727 .use_writepage = 1,
728 };
729
730 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
731 if (mpd.bio) {
732 int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
733 REQ_SYNC : 0);
734 mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
735 }
736 }
737 blk_finish_plug(&plug);
738 return ret;
739}
740EXPORT_SYMBOL(mpage_writepages);
741
742int mpage_writepage(struct page *page, get_block_t get_block,
743 struct writeback_control *wbc)
744{
745 struct mpage_data mpd = {
746 .bio = NULL,
747 .last_block_in_bio = 0,
748 .get_block = get_block,
749 .use_writepage = 0,
750 };
751 int ret = __mpage_writepage(page, wbc, &mpd);
752 if (mpd.bio) {
753 int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
754 REQ_SYNC : 0);
755 mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
756 }
757 return ret;
758}
759EXPORT_SYMBOL(mpage_writepage);
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_end_io(struct bio *bio)
47{
48 struct bio_vec *bv;
49 struct bvec_iter_all iter_all;
50
51 bio_for_each_segment_all(bv, bio, iter_all) {
52 struct page *page = bv->bv_page;
53 page_endio(page, bio_op(bio),
54 blk_status_to_errno(bio->bi_status));
55 }
56
57 bio_put(bio);
58}
59
60static struct bio *mpage_bio_submit(struct bio *bio)
61{
62 bio->bi_end_io = mpage_end_io;
63 guard_bio_eod(bio);
64 submit_bio(bio);
65 return NULL;
66}
67
68/*
69 * support function for mpage_readahead. The fs supplied get_block might
70 * return an up to date buffer. This is used to map that buffer into
71 * the page, which allows read_folio to avoid triggering a duplicate call
72 * to get_block.
73 *
74 * The idea is to avoid adding buffers to pages that don't already have
75 * them. So when the buffer is up to date and the page size == block size,
76 * this marks the page up to date instead of adding new buffers.
77 */
78static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
79 int page_block)
80{
81 struct inode *inode = folio->mapping->host;
82 struct buffer_head *page_bh, *head;
83 int block = 0;
84
85 head = folio_buffers(folio);
86 if (!head) {
87 /*
88 * don't make any buffers if there is only one buffer on
89 * the folio and the folio just needs to be set up to date
90 */
91 if (inode->i_blkbits == PAGE_SHIFT &&
92 buffer_uptodate(bh)) {
93 folio_mark_uptodate(folio);
94 return;
95 }
96 create_empty_buffers(&folio->page, i_blocksize(inode), 0);
97 head = folio_buffers(folio);
98 }
99
100 page_bh = head;
101 do {
102 if (block == page_block) {
103 page_bh->b_state = bh->b_state;
104 page_bh->b_bdev = bh->b_bdev;
105 page_bh->b_blocknr = bh->b_blocknr;
106 break;
107 }
108 page_bh = page_bh->b_this_page;
109 block++;
110 } while (page_bh != head);
111}
112
113struct mpage_readpage_args {
114 struct bio *bio;
115 struct folio *folio;
116 unsigned int nr_pages;
117 bool is_readahead;
118 sector_t last_block_in_bio;
119 struct buffer_head map_bh;
120 unsigned long first_logical_block;
121 get_block_t *get_block;
122};
123
124/*
125 * This is the worker routine which does all the work of mapping the disk
126 * blocks and constructs largest possible bios, submits them for IO if the
127 * blocks are not contiguous on the disk.
128 *
129 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
130 * represent the validity of its disk mapping and to decide when to do the next
131 * get_block() call.
132 */
133static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
134{
135 struct folio *folio = args->folio;
136 struct inode *inode = folio->mapping->host;
137 const unsigned blkbits = inode->i_blkbits;
138 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
139 const unsigned blocksize = 1 << blkbits;
140 struct buffer_head *map_bh = &args->map_bh;
141 sector_t block_in_file;
142 sector_t last_block;
143 sector_t last_block_in_file;
144 sector_t blocks[MAX_BUF_PER_PAGE];
145 unsigned page_block;
146 unsigned first_hole = blocks_per_page;
147 struct block_device *bdev = NULL;
148 int length;
149 int fully_mapped = 1;
150 blk_opf_t opf = REQ_OP_READ;
151 unsigned nblocks;
152 unsigned relative_block;
153 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
154
155 /* MAX_BUF_PER_PAGE, for example */
156 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
157
158 if (args->is_readahead) {
159 opf |= REQ_RAHEAD;
160 gfp |= __GFP_NORETRY | __GFP_NOWARN;
161 }
162
163 if (folio_buffers(folio))
164 goto confused;
165
166 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
167 last_block = block_in_file + args->nr_pages * blocks_per_page;
168 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
169 if (last_block > last_block_in_file)
170 last_block = last_block_in_file;
171 page_block = 0;
172
173 /*
174 * Map blocks using the result from the previous get_blocks call first.
175 */
176 nblocks = map_bh->b_size >> blkbits;
177 if (buffer_mapped(map_bh) &&
178 block_in_file > args->first_logical_block &&
179 block_in_file < (args->first_logical_block + nblocks)) {
180 unsigned map_offset = block_in_file - args->first_logical_block;
181 unsigned last = nblocks - map_offset;
182
183 for (relative_block = 0; ; relative_block++) {
184 if (relative_block == last) {
185 clear_buffer_mapped(map_bh);
186 break;
187 }
188 if (page_block == blocks_per_page)
189 break;
190 blocks[page_block] = map_bh->b_blocknr + map_offset +
191 relative_block;
192 page_block++;
193 block_in_file++;
194 }
195 bdev = map_bh->b_bdev;
196 }
197
198 /*
199 * Then do more get_blocks calls until we are done with this folio.
200 */
201 map_bh->b_page = &folio->page;
202 while (page_block < blocks_per_page) {
203 map_bh->b_state = 0;
204 map_bh->b_size = 0;
205
206 if (block_in_file < last_block) {
207 map_bh->b_size = (last_block-block_in_file) << blkbits;
208 if (args->get_block(inode, block_in_file, map_bh, 0))
209 goto confused;
210 args->first_logical_block = block_in_file;
211 }
212
213 if (!buffer_mapped(map_bh)) {
214 fully_mapped = 0;
215 if (first_hole == blocks_per_page)
216 first_hole = page_block;
217 page_block++;
218 block_in_file++;
219 continue;
220 }
221
222 /* some filesystems will copy data into the page during
223 * the get_block call, in which case we don't want to
224 * read it again. map_buffer_to_folio copies the data
225 * we just collected from get_block into the folio's buffers
226 * so read_folio doesn't have to repeat the get_block call
227 */
228 if (buffer_uptodate(map_bh)) {
229 map_buffer_to_folio(folio, map_bh, page_block);
230 goto confused;
231 }
232
233 if (first_hole != blocks_per_page)
234 goto confused; /* hole -> non-hole */
235
236 /* Contiguous blocks? */
237 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
238 goto confused;
239 nblocks = map_bh->b_size >> blkbits;
240 for (relative_block = 0; ; relative_block++) {
241 if (relative_block == nblocks) {
242 clear_buffer_mapped(map_bh);
243 break;
244 } else if (page_block == blocks_per_page)
245 break;
246 blocks[page_block] = map_bh->b_blocknr+relative_block;
247 page_block++;
248 block_in_file++;
249 }
250 bdev = map_bh->b_bdev;
251 }
252
253 if (first_hole != blocks_per_page) {
254 folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
255 if (first_hole == 0) {
256 folio_mark_uptodate(folio);
257 folio_unlock(folio);
258 goto out;
259 }
260 } else if (fully_mapped) {
261 folio_set_mappedtodisk(folio);
262 }
263
264 /*
265 * This folio will go to BIO. Do we need to send this BIO off first?
266 */
267 if (args->bio && (args->last_block_in_bio != blocks[0] - 1))
268 args->bio = mpage_bio_submit(args->bio);
269
270alloc_new:
271 if (args->bio == NULL) {
272 if (first_hole == blocks_per_page) {
273 if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
274 &folio->page))
275 goto out;
276 }
277 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
278 gfp);
279 if (args->bio == NULL)
280 goto confused;
281 args->bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
282 }
283
284 length = first_hole << blkbits;
285 if (!bio_add_folio(args->bio, folio, length, 0)) {
286 args->bio = mpage_bio_submit(args->bio);
287 goto alloc_new;
288 }
289
290 relative_block = block_in_file - args->first_logical_block;
291 nblocks = map_bh->b_size >> blkbits;
292 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
293 (first_hole != blocks_per_page))
294 args->bio = mpage_bio_submit(args->bio);
295 else
296 args->last_block_in_bio = blocks[blocks_per_page - 1];
297out:
298 return args->bio;
299
300confused:
301 if (args->bio)
302 args->bio = mpage_bio_submit(args->bio);
303 if (!folio_test_uptodate(folio))
304 block_read_full_folio(folio, args->get_block);
305 else
306 folio_unlock(folio);
307 goto out;
308}
309
310/**
311 * mpage_readahead - start reads against pages
312 * @rac: Describes which pages to read.
313 * @get_block: The filesystem's block mapper function.
314 *
315 * This function walks the pages and the blocks within each page, building and
316 * emitting large BIOs.
317 *
318 * If anything unusual happens, such as:
319 *
320 * - encountering a page which has buffers
321 * - encountering a page which has a non-hole after a hole
322 * - encountering a page with non-contiguous blocks
323 *
324 * then this code just gives up and calls the buffer_head-based read function.
325 * It does handle a page which has holes at the end - that is a common case:
326 * the end-of-file on blocksize < PAGE_SIZE setups.
327 *
328 * BH_Boundary explanation:
329 *
330 * There is a problem. The mpage read code assembles several pages, gets all
331 * their disk mappings, and then submits them all. That's fine, but obtaining
332 * the disk mappings may require I/O. Reads of indirect blocks, for example.
333 *
334 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
335 * submitted in the following order:
336 *
337 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
338 *
339 * because the indirect block has to be read to get the mappings of blocks
340 * 13,14,15,16. Obviously, this impacts performance.
341 *
342 * So what we do it to allow the filesystem's get_block() function to set
343 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
344 * after this one will require I/O against a block which is probably close to
345 * this one. So you should push what I/O you have currently accumulated.
346 *
347 * This all causes the disk requests to be issued in the correct order.
348 */
349void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
350{
351 struct folio *folio;
352 struct mpage_readpage_args args = {
353 .get_block = get_block,
354 .is_readahead = true,
355 };
356
357 while ((folio = readahead_folio(rac))) {
358 prefetchw(&folio->flags);
359 args.folio = folio;
360 args.nr_pages = readahead_count(rac);
361 args.bio = do_mpage_readpage(&args);
362 }
363 if (args.bio)
364 mpage_bio_submit(args.bio);
365}
366EXPORT_SYMBOL(mpage_readahead);
367
368/*
369 * This isn't called much at all
370 */
371int mpage_read_folio(struct folio *folio, get_block_t get_block)
372{
373 struct mpage_readpage_args args = {
374 .folio = folio,
375 .nr_pages = 1,
376 .get_block = get_block,
377 };
378
379 args.bio = do_mpage_readpage(&args);
380 if (args.bio)
381 mpage_bio_submit(args.bio);
382 return 0;
383}
384EXPORT_SYMBOL(mpage_read_folio);
385
386/*
387 * Writing is not so simple.
388 *
389 * If the page has buffers then they will be used for obtaining the disk
390 * mapping. We only support pages which are fully mapped-and-dirty, with a
391 * special case for pages which are unmapped at the end: end-of-file.
392 *
393 * If the page has no buffers (preferred) then the page is mapped here.
394 *
395 * If all blocks are found to be contiguous then the page can go into the
396 * BIO. Otherwise fall back to the mapping's writepage().
397 *
398 * FIXME: This code wants an estimate of how many pages are still to be
399 * written, so it can intelligently allocate a suitably-sized BIO. For now,
400 * just allocate full-size (16-page) BIOs.
401 */
402
403struct mpage_data {
404 struct bio *bio;
405 sector_t last_block_in_bio;
406 get_block_t *get_block;
407};
408
409/*
410 * We have our BIO, so we can now mark the buffers clean. Make
411 * sure to only clean buffers which we know we'll be writing.
412 */
413static void clean_buffers(struct page *page, unsigned first_unmapped)
414{
415 unsigned buffer_counter = 0;
416 struct buffer_head *bh, *head;
417 if (!page_has_buffers(page))
418 return;
419 head = page_buffers(page);
420 bh = head;
421
422 do {
423 if (buffer_counter++ == first_unmapped)
424 break;
425 clear_buffer_dirty(bh);
426 bh = bh->b_this_page;
427 } while (bh != head);
428
429 /*
430 * we cannot drop the bh if the page is not uptodate or a concurrent
431 * read_folio would fail to serialize with the bh and it would read from
432 * disk before we reach the platter.
433 */
434 if (buffer_heads_over_limit && PageUptodate(page))
435 try_to_free_buffers(page_folio(page));
436}
437
438/*
439 * For situations where we want to clean all buffers attached to a page.
440 * We don't need to calculate how many buffers are attached to the page,
441 * we just need to specify a number larger than the maximum number of buffers.
442 */
443void clean_page_buffers(struct page *page)
444{
445 clean_buffers(page, ~0U);
446}
447
448static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
449 void *data)
450{
451 struct mpage_data *mpd = data;
452 struct bio *bio = mpd->bio;
453 struct address_space *mapping = page->mapping;
454 struct inode *inode = page->mapping->host;
455 const unsigned blkbits = inode->i_blkbits;
456 unsigned long end_index;
457 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
458 sector_t last_block;
459 sector_t block_in_file;
460 sector_t blocks[MAX_BUF_PER_PAGE];
461 unsigned page_block;
462 unsigned first_unmapped = blocks_per_page;
463 struct block_device *bdev = NULL;
464 int boundary = 0;
465 sector_t boundary_block = 0;
466 struct block_device *boundary_bdev = NULL;
467 int length;
468 struct buffer_head map_bh;
469 loff_t i_size = i_size_read(inode);
470 int ret = 0;
471
472 if (page_has_buffers(page)) {
473 struct buffer_head *head = page_buffers(page);
474 struct buffer_head *bh = head;
475
476 /* If they're all mapped and dirty, do it */
477 page_block = 0;
478 do {
479 BUG_ON(buffer_locked(bh));
480 if (!buffer_mapped(bh)) {
481 /*
482 * unmapped dirty buffers are created by
483 * block_dirty_folio -> mmapped data
484 */
485 if (buffer_dirty(bh))
486 goto confused;
487 if (first_unmapped == blocks_per_page)
488 first_unmapped = page_block;
489 continue;
490 }
491
492 if (first_unmapped != blocks_per_page)
493 goto confused; /* hole -> non-hole */
494
495 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
496 goto confused;
497 if (page_block) {
498 if (bh->b_blocknr != blocks[page_block-1] + 1)
499 goto confused;
500 }
501 blocks[page_block++] = bh->b_blocknr;
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(!PageUptodate(page));
526 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
527 last_block = (i_size - 1) >> blkbits;
528 map_bh.b_page = page;
529 for (page_block = 0; page_block < blocks_per_page; ) {
530
531 map_bh.b_state = 0;
532 map_bh.b_size = 1 << blkbits;
533 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
534 goto confused;
535 if (buffer_new(&map_bh))
536 clean_bdev_bh_alias(&map_bh);
537 if (buffer_boundary(&map_bh)) {
538 boundary_block = map_bh.b_blocknr;
539 boundary_bdev = map_bh.b_bdev;
540 }
541 if (page_block) {
542 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
543 goto confused;
544 }
545 blocks[page_block++] = map_bh.b_blocknr;
546 boundary = buffer_boundary(&map_bh);
547 bdev = map_bh.b_bdev;
548 if (block_in_file == last_block)
549 break;
550 block_in_file++;
551 }
552 BUG_ON(page_block == 0);
553
554 first_unmapped = page_block;
555
556page_is_mapped:
557 end_index = i_size >> PAGE_SHIFT;
558 if (page->index >= end_index) {
559 /*
560 * The page straddles i_size. It must be zeroed out on each
561 * and every writepage invocation because it may be mmapped.
562 * "A file is mapped in multiples of the page size. For a file
563 * that is not a multiple of the page size, the remaining memory
564 * is zeroed when mapped, and writes to that region are not
565 * written out to the file."
566 */
567 unsigned offset = i_size & (PAGE_SIZE - 1);
568
569 if (page->index > end_index || !offset)
570 goto confused;
571 zero_user_segment(page, offset, PAGE_SIZE);
572 }
573
574 /*
575 * This page will go to BIO. Do we need to send this BIO off first?
576 */
577 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
578 bio = mpage_bio_submit(bio);
579
580alloc_new:
581 if (bio == NULL) {
582 if (first_unmapped == blocks_per_page) {
583 if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
584 page, wbc))
585 goto out;
586 }
587 bio = bio_alloc(bdev, BIO_MAX_VECS,
588 REQ_OP_WRITE | wbc_to_write_flags(wbc),
589 GFP_NOFS);
590 bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
591 wbc_init_bio(wbc, bio);
592 }
593
594 /*
595 * Must try to add the page before marking the buffer clean or
596 * the confused fail path above (OOM) will be very confused when
597 * it finds all bh marked clean (i.e. it will not write anything)
598 */
599 wbc_account_cgroup_owner(wbc, page, PAGE_SIZE);
600 length = first_unmapped << blkbits;
601 if (bio_add_page(bio, page, length, 0) < length) {
602 bio = mpage_bio_submit(bio);
603 goto alloc_new;
604 }
605
606 clean_buffers(page, first_unmapped);
607
608 BUG_ON(PageWriteback(page));
609 set_page_writeback(page);
610 unlock_page(page);
611 if (boundary || (first_unmapped != blocks_per_page)) {
612 bio = mpage_bio_submit(bio);
613 if (boundary_block) {
614 write_boundary_block(boundary_bdev,
615 boundary_block, 1 << blkbits);
616 }
617 } else {
618 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
619 }
620 goto out;
621
622confused:
623 if (bio)
624 bio = mpage_bio_submit(bio);
625
626 /*
627 * The caller has a ref on the inode, so *mapping is stable
628 */
629 ret = block_write_full_page(page, mpd->get_block, wbc);
630 mapping_set_error(mapping, ret);
631out:
632 mpd->bio = bio;
633 return ret;
634}
635
636/**
637 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
638 * @mapping: address space structure to write
639 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
640 * @get_block: the filesystem's block mapper function.
641 *
642 * This is a library function, which implements the writepages()
643 * address_space_operation.
644 *
645 * If a page is already under I/O, generic_writepages() skips it, even
646 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
647 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
648 * and msync() need to guarantee that all the data which was dirty at the time
649 * the call was made get new I/O started against them. If wbc->sync_mode is
650 * WB_SYNC_ALL then we were called for data integrity and we must wait for
651 * existing IO to complete.
652 */
653int
654mpage_writepages(struct address_space *mapping,
655 struct writeback_control *wbc, get_block_t get_block)
656{
657 struct mpage_data mpd = {
658 .get_block = get_block,
659 };
660 struct blk_plug plug;
661 int ret;
662
663 blk_start_plug(&plug);
664 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
665 if (mpd.bio)
666 mpage_bio_submit(mpd.bio);
667 blk_finish_plug(&plug);
668 return ret;
669}
670EXPORT_SYMBOL(mpage_writepages);