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 "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(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
62static 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
78static 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
86static 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 */
104static 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 head = create_empty_buffers(folio, i_blocksize(inode), 0);
123 }
124
125 page_bh = head;
126 do {
127 if (block == page_block) {
128 page_bh->b_state = bh->b_state;
129 page_bh->b_bdev = bh->b_bdev;
130 page_bh->b_blocknr = bh->b_blocknr;
131 break;
132 }
133 page_bh = page_bh->b_this_page;
134 block++;
135 } while (page_bh != head);
136}
137
138struct mpage_readpage_args {
139 struct bio *bio;
140 struct folio *folio;
141 unsigned int nr_pages;
142 bool is_readahead;
143 sector_t last_block_in_bio;
144 struct buffer_head map_bh;
145 unsigned long first_logical_block;
146 get_block_t *get_block;
147};
148
149/*
150 * This is the worker routine which does all the work of mapping the disk
151 * blocks and constructs largest possible bios, submits them for IO if the
152 * blocks are not contiguous on the disk.
153 *
154 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
155 * represent the validity of its disk mapping and to decide when to do the next
156 * get_block() call.
157 */
158static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
159{
160 struct folio *folio = args->folio;
161 struct inode *inode = folio->mapping->host;
162 const unsigned blkbits = inode->i_blkbits;
163 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
164 const unsigned blocksize = 1 << blkbits;
165 struct buffer_head *map_bh = &args->map_bh;
166 sector_t block_in_file;
167 sector_t last_block;
168 sector_t last_block_in_file;
169 sector_t first_block;
170 unsigned page_block;
171 unsigned first_hole = blocks_per_page;
172 struct block_device *bdev = NULL;
173 int length;
174 int fully_mapped = 1;
175 blk_opf_t opf = REQ_OP_READ;
176 unsigned nblocks;
177 unsigned relative_block;
178 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
179
180 /* MAX_BUF_PER_PAGE, for example */
181 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
182
183 if (args->is_readahead) {
184 opf |= REQ_RAHEAD;
185 gfp |= __GFP_NORETRY | __GFP_NOWARN;
186 }
187
188 if (folio_buffers(folio))
189 goto confused;
190
191 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
192 last_block = block_in_file + args->nr_pages * blocks_per_page;
193 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
194 if (last_block > last_block_in_file)
195 last_block = last_block_in_file;
196 page_block = 0;
197
198 /*
199 * Map blocks using the result from the previous get_blocks call first.
200 */
201 nblocks = map_bh->b_size >> blkbits;
202 if (buffer_mapped(map_bh) &&
203 block_in_file > args->first_logical_block &&
204 block_in_file < (args->first_logical_block + nblocks)) {
205 unsigned map_offset = block_in_file - args->first_logical_block;
206 unsigned last = nblocks - map_offset;
207
208 first_block = map_bh->b_blocknr + map_offset;
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 page_block++;
217 block_in_file++;
218 }
219 bdev = map_bh->b_bdev;
220 }
221
222 /*
223 * Then do more get_blocks calls until we are done with this folio.
224 */
225 map_bh->b_folio = folio;
226 while (page_block < blocks_per_page) {
227 map_bh->b_state = 0;
228 map_bh->b_size = 0;
229
230 if (block_in_file < last_block) {
231 map_bh->b_size = (last_block-block_in_file) << blkbits;
232 if (args->get_block(inode, block_in_file, map_bh, 0))
233 goto confused;
234 args->first_logical_block = block_in_file;
235 }
236
237 if (!buffer_mapped(map_bh)) {
238 fully_mapped = 0;
239 if (first_hole == blocks_per_page)
240 first_hole = page_block;
241 page_block++;
242 block_in_file++;
243 continue;
244 }
245
246 /* some filesystems will copy data into the page during
247 * the get_block call, in which case we don't want to
248 * read it again. map_buffer_to_folio copies the data
249 * we just collected from get_block into the folio's buffers
250 * so read_folio doesn't have to repeat the get_block call
251 */
252 if (buffer_uptodate(map_bh)) {
253 map_buffer_to_folio(folio, map_bh, page_block);
254 goto confused;
255 }
256
257 if (first_hole != blocks_per_page)
258 goto confused; /* hole -> non-hole */
259
260 /* Contiguous blocks? */
261 if (!page_block)
262 first_block = map_bh->b_blocknr;
263 else if (first_block + page_block != map_bh->b_blocknr)
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 page_block++;
273 block_in_file++;
274 }
275 bdev = map_bh->b_bdev;
276 }
277
278 if (first_hole != blocks_per_page) {
279 folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
280 if (first_hole == 0) {
281 folio_mark_uptodate(folio);
282 folio_unlock(folio);
283 goto out;
284 }
285 } else if (fully_mapped) {
286 folio_set_mappedtodisk(folio);
287 }
288
289 /*
290 * This folio will go to BIO. Do we need to send this BIO off first?
291 */
292 if (args->bio && (args->last_block_in_bio != first_block - 1))
293 args->bio = mpage_bio_submit_read(args->bio);
294
295alloc_new:
296 if (args->bio == NULL) {
297 args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
298 gfp);
299 if (args->bio == NULL)
300 goto confused;
301 args->bio->bi_iter.bi_sector = first_block << (blkbits - 9);
302 }
303
304 length = first_hole << blkbits;
305 if (!bio_add_folio(args->bio, folio, length, 0)) {
306 args->bio = mpage_bio_submit_read(args->bio);
307 goto alloc_new;
308 }
309
310 relative_block = block_in_file - args->first_logical_block;
311 nblocks = map_bh->b_size >> blkbits;
312 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
313 (first_hole != blocks_per_page))
314 args->bio = mpage_bio_submit_read(args->bio);
315 else
316 args->last_block_in_bio = first_block + blocks_per_page - 1;
317out:
318 return args->bio;
319
320confused:
321 if (args->bio)
322 args->bio = mpage_bio_submit_read(args->bio);
323 if (!folio_test_uptodate(folio))
324 block_read_full_folio(folio, args->get_block);
325 else
326 folio_unlock(folio);
327 goto out;
328}
329
330/**
331 * mpage_readahead - start reads against pages
332 * @rac: Describes which pages to read.
333 * @get_block: The filesystem's block mapper function.
334 *
335 * This function walks the pages and the blocks within each page, building and
336 * emitting large BIOs.
337 *
338 * If anything unusual happens, such as:
339 *
340 * - encountering a page which has buffers
341 * - encountering a page which has a non-hole after a hole
342 * - encountering a page with non-contiguous blocks
343 *
344 * then this code just gives up and calls the buffer_head-based read function.
345 * It does handle a page which has holes at the end - that is a common case:
346 * the end-of-file on blocksize < PAGE_SIZE setups.
347 *
348 * BH_Boundary explanation:
349 *
350 * There is a problem. The mpage read code assembles several pages, gets all
351 * their disk mappings, and then submits them all. That's fine, but obtaining
352 * the disk mappings may require I/O. Reads of indirect blocks, for example.
353 *
354 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
355 * submitted in the following order:
356 *
357 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
358 *
359 * because the indirect block has to be read to get the mappings of blocks
360 * 13,14,15,16. Obviously, this impacts performance.
361 *
362 * So what we do it to allow the filesystem's get_block() function to set
363 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
364 * after this one will require I/O against a block which is probably close to
365 * this one. So you should push what I/O you have currently accumulated.
366 *
367 * This all causes the disk requests to be issued in the correct order.
368 */
369void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
370{
371 struct folio *folio;
372 struct mpage_readpage_args args = {
373 .get_block = get_block,
374 .is_readahead = true,
375 };
376
377 while ((folio = readahead_folio(rac))) {
378 prefetchw(&folio->flags);
379 args.folio = folio;
380 args.nr_pages = readahead_count(rac);
381 args.bio = do_mpage_readpage(&args);
382 }
383 if (args.bio)
384 mpage_bio_submit_read(args.bio);
385}
386EXPORT_SYMBOL(mpage_readahead);
387
388/*
389 * This isn't called much at all
390 */
391int mpage_read_folio(struct folio *folio, get_block_t get_block)
392{
393 struct mpage_readpage_args args = {
394 .folio = folio,
395 .nr_pages = 1,
396 .get_block = get_block,
397 };
398
399 args.bio = do_mpage_readpage(&args);
400 if (args.bio)
401 mpage_bio_submit_read(args.bio);
402 return 0;
403}
404EXPORT_SYMBOL(mpage_read_folio);
405
406/*
407 * Writing is not so simple.
408 *
409 * If the page has buffers then they will be used for obtaining the disk
410 * mapping. We only support pages which are fully mapped-and-dirty, with a
411 * special case for pages which are unmapped at the end: end-of-file.
412 *
413 * If the page has no buffers (preferred) then the page is mapped here.
414 *
415 * If all blocks are found to be contiguous then the page can go into the
416 * BIO. Otherwise fall back to the mapping's writepage().
417 *
418 * FIXME: This code wants an estimate of how many pages are still to be
419 * written, so it can intelligently allocate a suitably-sized BIO. For now,
420 * just allocate full-size (16-page) BIOs.
421 */
422
423struct mpage_data {
424 struct bio *bio;
425 sector_t last_block_in_bio;
426 get_block_t *get_block;
427};
428
429/*
430 * We have our BIO, so we can now mark the buffers clean. Make
431 * sure to only clean buffers which we know we'll be writing.
432 */
433static void clean_buffers(struct folio *folio, unsigned first_unmapped)
434{
435 unsigned buffer_counter = 0;
436 struct buffer_head *bh, *head = folio_buffers(folio);
437
438 if (!head)
439 return;
440 bh = head;
441
442 do {
443 if (buffer_counter++ == first_unmapped)
444 break;
445 clear_buffer_dirty(bh);
446 bh = bh->b_this_page;
447 } while (bh != head);
448
449 /*
450 * we cannot drop the bh if the page is not uptodate or a concurrent
451 * read_folio would fail to serialize with the bh and it would read from
452 * disk before we reach the platter.
453 */
454 if (buffer_heads_over_limit && folio_test_uptodate(folio))
455 try_to_free_buffers(folio);
456}
457
458static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
459 void *data)
460{
461 struct mpage_data *mpd = data;
462 struct bio *bio = mpd->bio;
463 struct address_space *mapping = folio->mapping;
464 struct inode *inode = mapping->host;
465 const unsigned blkbits = inode->i_blkbits;
466 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
467 sector_t last_block;
468 sector_t block_in_file;
469 sector_t first_block;
470 unsigned page_block;
471 unsigned first_unmapped = blocks_per_page;
472 struct block_device *bdev = NULL;
473 int boundary = 0;
474 sector_t boundary_block = 0;
475 struct block_device *boundary_bdev = NULL;
476 size_t length;
477 struct buffer_head map_bh;
478 loff_t i_size = i_size_read(inode);
479 int ret = 0;
480 struct buffer_head *head = folio_buffers(folio);
481
482 if (head) {
483 struct buffer_head *bh = head;
484
485 /* If they're all mapped and dirty, do it */
486 page_block = 0;
487 do {
488 BUG_ON(buffer_locked(bh));
489 if (!buffer_mapped(bh)) {
490 /*
491 * unmapped dirty buffers are created by
492 * block_dirty_folio -> mmapped data
493 */
494 if (buffer_dirty(bh))
495 goto confused;
496 if (first_unmapped == blocks_per_page)
497 first_unmapped = page_block;
498 continue;
499 }
500
501 if (first_unmapped != blocks_per_page)
502 goto confused; /* hole -> non-hole */
503
504 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
505 goto confused;
506 if (page_block) {
507 if (bh->b_blocknr != first_block + page_block)
508 goto confused;
509 } else {
510 first_block = bh->b_blocknr;
511 }
512 page_block++;
513 boundary = buffer_boundary(bh);
514 if (boundary) {
515 boundary_block = bh->b_blocknr;
516 boundary_bdev = bh->b_bdev;
517 }
518 bdev = bh->b_bdev;
519 } while ((bh = bh->b_this_page) != head);
520
521 if (first_unmapped)
522 goto page_is_mapped;
523
524 /*
525 * Page has buffers, but they are all unmapped. The page was
526 * created by pagein or read over a hole which was handled by
527 * block_read_full_folio(). If this address_space is also
528 * using mpage_readahead then this can rarely happen.
529 */
530 goto confused;
531 }
532
533 /*
534 * The page has no buffers: map it to disk
535 */
536 BUG_ON(!folio_test_uptodate(folio));
537 block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
538 /*
539 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
540 * space.
541 */
542 if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
543 goto page_is_mapped;
544 last_block = (i_size - 1) >> blkbits;
545 map_bh.b_folio = folio;
546 for (page_block = 0; page_block < blocks_per_page; ) {
547
548 map_bh.b_state = 0;
549 map_bh.b_size = 1 << blkbits;
550 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
551 goto confused;
552 if (!buffer_mapped(&map_bh))
553 goto confused;
554 if (buffer_new(&map_bh))
555 clean_bdev_bh_alias(&map_bh);
556 if (buffer_boundary(&map_bh)) {
557 boundary_block = map_bh.b_blocknr;
558 boundary_bdev = map_bh.b_bdev;
559 }
560 if (page_block) {
561 if (map_bh.b_blocknr != first_block + page_block)
562 goto confused;
563 } else {
564 first_block = map_bh.b_blocknr;
565 }
566 page_block++;
567 boundary = buffer_boundary(&map_bh);
568 bdev = map_bh.b_bdev;
569 if (block_in_file == last_block)
570 break;
571 block_in_file++;
572 }
573 BUG_ON(page_block == 0);
574
575 first_unmapped = page_block;
576
577page_is_mapped:
578 /* Don't bother writing beyond EOF, truncate will discard the folio */
579 if (folio_pos(folio) >= i_size)
580 goto confused;
581 length = folio_size(folio);
582 if (folio_pos(folio) + length > i_size) {
583 /*
584 * The page straddles i_size. It must be zeroed out on each
585 * and every writepage invocation because it may be mmapped.
586 * "A file is mapped in multiples of the page size. For a file
587 * that is not a multiple of the page size, the remaining memory
588 * is zeroed when mapped, and writes to that region are not
589 * written out to the file."
590 */
591 length = i_size - folio_pos(folio);
592 folio_zero_segment(folio, length, folio_size(folio));
593 }
594
595 /*
596 * This page will go to BIO. Do we need to send this BIO off first?
597 */
598 if (bio && mpd->last_block_in_bio != first_block - 1)
599 bio = mpage_bio_submit_write(bio);
600
601alloc_new:
602 if (bio == NULL) {
603 bio = bio_alloc(bdev, BIO_MAX_VECS,
604 REQ_OP_WRITE | wbc_to_write_flags(wbc),
605 GFP_NOFS);
606 bio->bi_iter.bi_sector = first_block << (blkbits - 9);
607 wbc_init_bio(wbc, bio);
608 bio->bi_write_hint = inode->i_write_hint;
609 }
610
611 /*
612 * Must try to add the page before marking the buffer clean or
613 * the confused fail path above (OOM) will be very confused when
614 * it finds all bh marked clean (i.e. it will not write anything)
615 */
616 wbc_account_cgroup_owner(wbc, &folio->page, folio_size(folio));
617 length = first_unmapped << blkbits;
618 if (!bio_add_folio(bio, folio, length, 0)) {
619 bio = mpage_bio_submit_write(bio);
620 goto alloc_new;
621 }
622
623 clean_buffers(folio, first_unmapped);
624
625 BUG_ON(folio_test_writeback(folio));
626 folio_start_writeback(folio);
627 folio_unlock(folio);
628 if (boundary || (first_unmapped != blocks_per_page)) {
629 bio = mpage_bio_submit_write(bio);
630 if (boundary_block) {
631 write_boundary_block(boundary_bdev,
632 boundary_block, 1 << blkbits);
633 }
634 } else {
635 mpd->last_block_in_bio = first_block + blocks_per_page - 1;
636 }
637 goto out;
638
639confused:
640 if (bio)
641 bio = mpage_bio_submit_write(bio);
642
643 /*
644 * The caller has a ref on the inode, so *mapping is stable
645 */
646 ret = block_write_full_folio(folio, wbc, mpd->get_block);
647 mapping_set_error(mapping, ret);
648out:
649 mpd->bio = bio;
650 return ret;
651}
652
653/**
654 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
655 * @mapping: address space structure to write
656 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
657 * @get_block: the filesystem's block mapper function.
658 *
659 * This is a library function, which implements the writepages()
660 * address_space_operation.
661 */
662int
663mpage_writepages(struct address_space *mapping,
664 struct writeback_control *wbc, get_block_t get_block)
665{
666 struct mpage_data mpd = {
667 .get_block = get_block,
668 };
669 struct blk_plug plug;
670 int ret;
671
672 blk_start_plug(&plug);
673 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
674 if (mpd.bio)
675 mpage_bio_submit_write(mpd.bio);
676 blk_finish_plug(&plug);
677 return ret;
678}
679EXPORT_SYMBOL(mpage_writepages);
1/*
2 * fs/mpage.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains functions related to preparing and submitting BIOs which contain
7 * multiple pagecache pages.
8 *
9 * 15May2002 Andrew Morton
10 * Initial version
11 * 27Jun2002 axboe@suse.de
12 * use bio_add_page() to build bio's just the right size
13 */
14
15#include <linux/kernel.h>
16#include <linux/export.h>
17#include <linux/mm.h>
18#include <linux/kdev_t.h>
19#include <linux/gfp.h>
20#include <linux/bio.h>
21#include <linux/fs.h>
22#include <linux/buffer_head.h>
23#include <linux/blkdev.h>
24#include <linux/highmem.h>
25#include <linux/prefetch.h>
26#include <linux/mpage.h>
27#include <linux/writeback.h>
28#include <linux/backing-dev.h>
29#include <linux/pagevec.h>
30#include <linux/cleancache.h>
31
32/*
33 * I/O completion handler for multipage BIOs.
34 *
35 * The mpage code never puts partial pages into a BIO (except for end-of-file).
36 * If a page does not map to a contiguous run of blocks then it simply falls
37 * back to block_read_full_page().
38 *
39 * Why is this? If a page's completion depends on a number of different BIOs
40 * which can complete in any order (or at the same time) then determining the
41 * status of that page is hard. See end_buffer_async_read() for the details.
42 * There is no point in duplicating all that complexity.
43 */
44static void mpage_end_io(struct bio *bio, int err)
45{
46 struct bio_vec *bv;
47 int i;
48
49 bio_for_each_segment_all(bv, bio, i) {
50 struct page *page = bv->bv_page;
51
52 if (bio_data_dir(bio) == READ) {
53 if (!err) {
54 SetPageUptodate(page);
55 } else {
56 ClearPageUptodate(page);
57 SetPageError(page);
58 }
59 unlock_page(page);
60 } else { /* bio_data_dir(bio) == WRITE */
61 if (err) {
62 SetPageError(page);
63 if (page->mapping)
64 set_bit(AS_EIO, &page->mapping->flags);
65 }
66 end_page_writeback(page);
67 }
68 }
69
70 bio_put(bio);
71}
72
73static struct bio *mpage_bio_submit(int rw, struct bio *bio)
74{
75 bio->bi_end_io = mpage_end_io;
76 submit_bio(rw, bio);
77 return NULL;
78}
79
80static struct bio *
81mpage_alloc(struct block_device *bdev,
82 sector_t first_sector, int nr_vecs,
83 gfp_t gfp_flags)
84{
85 struct bio *bio;
86
87 bio = bio_alloc(gfp_flags, nr_vecs);
88
89 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
90 while (!bio && (nr_vecs /= 2))
91 bio = bio_alloc(gfp_flags, nr_vecs);
92 }
93
94 if (bio) {
95 bio->bi_bdev = bdev;
96 bio->bi_iter.bi_sector = first_sector;
97 }
98 return bio;
99}
100
101/*
102 * support function for mpage_readpages. The fs supplied get_block might
103 * return an up to date buffer. This is used to map that buffer into
104 * the page, which allows readpage to avoid triggering a duplicate call
105 * to get_block.
106 *
107 * The idea is to avoid adding buffers to pages that don't already have
108 * them. So when the buffer is up to date and the page size == block size,
109 * this marks the page up to date instead of adding new buffers.
110 */
111static void
112map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
113{
114 struct inode *inode = page->mapping->host;
115 struct buffer_head *page_bh, *head;
116 int block = 0;
117
118 if (!page_has_buffers(page)) {
119 /*
120 * don't make any buffers if there is only one buffer on
121 * the page and the page just needs to be set up to date
122 */
123 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
124 buffer_uptodate(bh)) {
125 SetPageUptodate(page);
126 return;
127 }
128 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
129 }
130 head = page_buffers(page);
131 page_bh = head;
132 do {
133 if (block == page_block) {
134 page_bh->b_state = bh->b_state;
135 page_bh->b_bdev = bh->b_bdev;
136 page_bh->b_blocknr = bh->b_blocknr;
137 break;
138 }
139 page_bh = page_bh->b_this_page;
140 block++;
141 } while (page_bh != head);
142}
143
144/*
145 * This is the worker routine which does all the work of mapping the disk
146 * blocks and constructs largest possible bios, submits them for IO if the
147 * blocks are not contiguous on the disk.
148 *
149 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
150 * represent the validity of its disk mapping and to decide when to do the next
151 * get_block() call.
152 */
153static struct bio *
154do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
155 sector_t *last_block_in_bio, struct buffer_head *map_bh,
156 unsigned long *first_logical_block, get_block_t get_block)
157{
158 struct inode *inode = page->mapping->host;
159 const unsigned blkbits = inode->i_blkbits;
160 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
161 const unsigned blocksize = 1 << blkbits;
162 sector_t block_in_file;
163 sector_t last_block;
164 sector_t last_block_in_file;
165 sector_t blocks[MAX_BUF_PER_PAGE];
166 unsigned page_block;
167 unsigned first_hole = blocks_per_page;
168 struct block_device *bdev = NULL;
169 int length;
170 int fully_mapped = 1;
171 unsigned nblocks;
172 unsigned relative_block;
173
174 if (page_has_buffers(page))
175 goto confused;
176
177 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
178 last_block = block_in_file + nr_pages * blocks_per_page;
179 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
180 if (last_block > last_block_in_file)
181 last_block = last_block_in_file;
182 page_block = 0;
183
184 /*
185 * Map blocks using the result from the previous get_blocks call first.
186 */
187 nblocks = map_bh->b_size >> blkbits;
188 if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
189 block_in_file < (*first_logical_block + nblocks)) {
190 unsigned map_offset = block_in_file - *first_logical_block;
191 unsigned last = nblocks - map_offset;
192
193 for (relative_block = 0; ; relative_block++) {
194 if (relative_block == last) {
195 clear_buffer_mapped(map_bh);
196 break;
197 }
198 if (page_block == blocks_per_page)
199 break;
200 blocks[page_block] = map_bh->b_blocknr + map_offset +
201 relative_block;
202 page_block++;
203 block_in_file++;
204 }
205 bdev = map_bh->b_bdev;
206 }
207
208 /*
209 * Then do more get_blocks calls until we are done with this page.
210 */
211 map_bh->b_page = page;
212 while (page_block < blocks_per_page) {
213 map_bh->b_state = 0;
214 map_bh->b_size = 0;
215
216 if (block_in_file < last_block) {
217 map_bh->b_size = (last_block-block_in_file) << blkbits;
218 if (get_block(inode, block_in_file, map_bh, 0))
219 goto confused;
220 *first_logical_block = block_in_file;
221 }
222
223 if (!buffer_mapped(map_bh)) {
224 fully_mapped = 0;
225 if (first_hole == blocks_per_page)
226 first_hole = page_block;
227 page_block++;
228 block_in_file++;
229 continue;
230 }
231
232 /* some filesystems will copy data into the page during
233 * the get_block call, in which case we don't want to
234 * read it again. map_buffer_to_page copies the data
235 * we just collected from get_block into the page's buffers
236 * so readpage doesn't have to repeat the get_block call
237 */
238 if (buffer_uptodate(map_bh)) {
239 map_buffer_to_page(page, map_bh, page_block);
240 goto confused;
241 }
242
243 if (first_hole != blocks_per_page)
244 goto confused; /* hole -> non-hole */
245
246 /* Contiguous blocks? */
247 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
248 goto confused;
249 nblocks = map_bh->b_size >> blkbits;
250 for (relative_block = 0; ; relative_block++) {
251 if (relative_block == nblocks) {
252 clear_buffer_mapped(map_bh);
253 break;
254 } else if (page_block == blocks_per_page)
255 break;
256 blocks[page_block] = map_bh->b_blocknr+relative_block;
257 page_block++;
258 block_in_file++;
259 }
260 bdev = map_bh->b_bdev;
261 }
262
263 if (first_hole != blocks_per_page) {
264 zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
265 if (first_hole == 0) {
266 SetPageUptodate(page);
267 unlock_page(page);
268 goto out;
269 }
270 } else if (fully_mapped) {
271 SetPageMappedToDisk(page);
272 }
273
274 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
275 cleancache_get_page(page) == 0) {
276 SetPageUptodate(page);
277 goto confused;
278 }
279
280 /*
281 * This page will go to BIO. Do we need to send this BIO off first?
282 */
283 if (bio && (*last_block_in_bio != blocks[0] - 1))
284 bio = mpage_bio_submit(READ, bio);
285
286alloc_new:
287 if (bio == NULL) {
288 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
289 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
290 GFP_KERNEL);
291 if (bio == NULL)
292 goto confused;
293 }
294
295 length = first_hole << blkbits;
296 if (bio_add_page(bio, page, length, 0) < length) {
297 bio = mpage_bio_submit(READ, bio);
298 goto alloc_new;
299 }
300
301 relative_block = block_in_file - *first_logical_block;
302 nblocks = map_bh->b_size >> blkbits;
303 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
304 (first_hole != blocks_per_page))
305 bio = mpage_bio_submit(READ, bio);
306 else
307 *last_block_in_bio = blocks[blocks_per_page - 1];
308out:
309 return bio;
310
311confused:
312 if (bio)
313 bio = mpage_bio_submit(READ, bio);
314 if (!PageUptodate(page))
315 block_read_full_page(page, get_block);
316 else
317 unlock_page(page);
318 goto out;
319}
320
321/**
322 * mpage_readpages - populate an address space with some pages & start reads against them
323 * @mapping: the address_space
324 * @pages: The address of a list_head which contains the target pages. These
325 * pages have their ->index populated and are otherwise uninitialised.
326 * The page at @pages->prev has the lowest file offset, and reads should be
327 * issued in @pages->prev to @pages->next order.
328 * @nr_pages: The number of pages at *@pages
329 * @get_block: The filesystem's block mapper function.
330 *
331 * This function walks the pages and the blocks within each page, building and
332 * emitting large BIOs.
333 *
334 * If anything unusual happens, such as:
335 *
336 * - encountering a page which has buffers
337 * - encountering a page which has a non-hole after a hole
338 * - encountering a page with non-contiguous blocks
339 *
340 * then this code just gives up and calls the buffer_head-based read function.
341 * It does handle a page which has holes at the end - that is a common case:
342 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
343 *
344 * BH_Boundary explanation:
345 *
346 * There is a problem. The mpage read code assembles several pages, gets all
347 * their disk mappings, and then submits them all. That's fine, but obtaining
348 * the disk mappings may require I/O. Reads of indirect blocks, for example.
349 *
350 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
351 * submitted in the following order:
352 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
353 *
354 * because the indirect block has to be read to get the mappings of blocks
355 * 13,14,15,16. Obviously, this impacts performance.
356 *
357 * So what we do it to allow the filesystem's get_block() function to set
358 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
359 * after this one will require I/O against a block which is probably close to
360 * this one. So you should push what I/O you have currently accumulated.
361 *
362 * This all causes the disk requests to be issued in the correct order.
363 */
364int
365mpage_readpages(struct address_space *mapping, struct list_head *pages,
366 unsigned nr_pages, get_block_t get_block)
367{
368 struct bio *bio = NULL;
369 unsigned page_idx;
370 sector_t last_block_in_bio = 0;
371 struct buffer_head map_bh;
372 unsigned long first_logical_block = 0;
373
374 map_bh.b_state = 0;
375 map_bh.b_size = 0;
376 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
377 struct page *page = list_entry(pages->prev, struct page, lru);
378
379 prefetchw(&page->flags);
380 list_del(&page->lru);
381 if (!add_to_page_cache_lru(page, mapping,
382 page->index, GFP_KERNEL)) {
383 bio = do_mpage_readpage(bio, page,
384 nr_pages - page_idx,
385 &last_block_in_bio, &map_bh,
386 &first_logical_block,
387 get_block);
388 }
389 page_cache_release(page);
390 }
391 BUG_ON(!list_empty(pages));
392 if (bio)
393 mpage_bio_submit(READ, bio);
394 return 0;
395}
396EXPORT_SYMBOL(mpage_readpages);
397
398/*
399 * This isn't called much at all
400 */
401int mpage_readpage(struct page *page, get_block_t get_block)
402{
403 struct bio *bio = NULL;
404 sector_t last_block_in_bio = 0;
405 struct buffer_head map_bh;
406 unsigned long first_logical_block = 0;
407
408 map_bh.b_state = 0;
409 map_bh.b_size = 0;
410 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
411 &map_bh, &first_logical_block, get_block);
412 if (bio)
413 mpage_bio_submit(READ, bio);
414 return 0;
415}
416EXPORT_SYMBOL(mpage_readpage);
417
418/*
419 * Writing is not so simple.
420 *
421 * If the page has buffers then they will be used for obtaining the disk
422 * mapping. We only support pages which are fully mapped-and-dirty, with a
423 * special case for pages which are unmapped at the end: end-of-file.
424 *
425 * If the page has no buffers (preferred) then the page is mapped here.
426 *
427 * If all blocks are found to be contiguous then the page can go into the
428 * BIO. Otherwise fall back to the mapping's writepage().
429 *
430 * FIXME: This code wants an estimate of how many pages are still to be
431 * written, so it can intelligently allocate a suitably-sized BIO. For now,
432 * just allocate full-size (16-page) BIOs.
433 */
434
435struct mpage_data {
436 struct bio *bio;
437 sector_t last_block_in_bio;
438 get_block_t *get_block;
439 unsigned use_writepage;
440};
441
442static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
443 void *data)
444{
445 struct mpage_data *mpd = data;
446 struct bio *bio = mpd->bio;
447 struct address_space *mapping = page->mapping;
448 struct inode *inode = page->mapping->host;
449 const unsigned blkbits = inode->i_blkbits;
450 unsigned long end_index;
451 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
452 sector_t last_block;
453 sector_t block_in_file;
454 sector_t blocks[MAX_BUF_PER_PAGE];
455 unsigned page_block;
456 unsigned first_unmapped = blocks_per_page;
457 struct block_device *bdev = NULL;
458 int boundary = 0;
459 sector_t boundary_block = 0;
460 struct block_device *boundary_bdev = NULL;
461 int length;
462 struct buffer_head map_bh;
463 loff_t i_size = i_size_read(inode);
464 int ret = 0;
465
466 if (page_has_buffers(page)) {
467 struct buffer_head *head = page_buffers(page);
468 struct buffer_head *bh = head;
469
470 /* If they're all mapped and dirty, do it */
471 page_block = 0;
472 do {
473 BUG_ON(buffer_locked(bh));
474 if (!buffer_mapped(bh)) {
475 /*
476 * unmapped dirty buffers are created by
477 * __set_page_dirty_buffers -> mmapped data
478 */
479 if (buffer_dirty(bh))
480 goto confused;
481 if (first_unmapped == blocks_per_page)
482 first_unmapped = page_block;
483 continue;
484 }
485
486 if (first_unmapped != blocks_per_page)
487 goto confused; /* hole -> non-hole */
488
489 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
490 goto confused;
491 if (page_block) {
492 if (bh->b_blocknr != blocks[page_block-1] + 1)
493 goto confused;
494 }
495 blocks[page_block++] = bh->b_blocknr;
496 boundary = buffer_boundary(bh);
497 if (boundary) {
498 boundary_block = bh->b_blocknr;
499 boundary_bdev = bh->b_bdev;
500 }
501 bdev = bh->b_bdev;
502 } while ((bh = bh->b_this_page) != head);
503
504 if (first_unmapped)
505 goto page_is_mapped;
506
507 /*
508 * Page has buffers, but they are all unmapped. The page was
509 * created by pagein or read over a hole which was handled by
510 * block_read_full_page(). If this address_space is also
511 * using mpage_readpages then this can rarely happen.
512 */
513 goto confused;
514 }
515
516 /*
517 * The page has no buffers: map it to disk
518 */
519 BUG_ON(!PageUptodate(page));
520 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
521 last_block = (i_size - 1) >> blkbits;
522 map_bh.b_page = page;
523 for (page_block = 0; page_block < blocks_per_page; ) {
524
525 map_bh.b_state = 0;
526 map_bh.b_size = 1 << blkbits;
527 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
528 goto confused;
529 if (buffer_new(&map_bh))
530 unmap_underlying_metadata(map_bh.b_bdev,
531 map_bh.b_blocknr);
532 if (buffer_boundary(&map_bh)) {
533 boundary_block = map_bh.b_blocknr;
534 boundary_bdev = map_bh.b_bdev;
535 }
536 if (page_block) {
537 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
538 goto confused;
539 }
540 blocks[page_block++] = map_bh.b_blocknr;
541 boundary = buffer_boundary(&map_bh);
542 bdev = map_bh.b_bdev;
543 if (block_in_file == last_block)
544 break;
545 block_in_file++;
546 }
547 BUG_ON(page_block == 0);
548
549 first_unmapped = page_block;
550
551page_is_mapped:
552 end_index = i_size >> PAGE_CACHE_SHIFT;
553 if (page->index >= end_index) {
554 /*
555 * The page straddles i_size. It must be zeroed out on each
556 * and every writepage invocation because it may be mmapped.
557 * "A file is mapped in multiples of the page size. For a file
558 * that is not a multiple of the page size, the remaining memory
559 * is zeroed when mapped, and writes to that region are not
560 * written out to the file."
561 */
562 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
563
564 if (page->index > end_index || !offset)
565 goto confused;
566 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
567 }
568
569 /*
570 * This page will go to BIO. Do we need to send this BIO off first?
571 */
572 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
573 bio = mpage_bio_submit(WRITE, bio);
574
575alloc_new:
576 if (bio == NULL) {
577 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
578 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
579 if (bio == NULL)
580 goto confused;
581 }
582
583 /*
584 * Must try to add the page before marking the buffer clean or
585 * the confused fail path above (OOM) will be very confused when
586 * it finds all bh marked clean (i.e. it will not write anything)
587 */
588 length = first_unmapped << blkbits;
589 if (bio_add_page(bio, page, length, 0) < length) {
590 bio = mpage_bio_submit(WRITE, bio);
591 goto alloc_new;
592 }
593
594 /*
595 * OK, we have our BIO, so we can now mark the buffers clean. Make
596 * sure to only clean buffers which we know we'll be writing.
597 */
598 if (page_has_buffers(page)) {
599 struct buffer_head *head = page_buffers(page);
600 struct buffer_head *bh = head;
601 unsigned buffer_counter = 0;
602
603 do {
604 if (buffer_counter++ == first_unmapped)
605 break;
606 clear_buffer_dirty(bh);
607 bh = bh->b_this_page;
608 } while (bh != head);
609
610 /*
611 * we cannot drop the bh if the page is not uptodate
612 * or a concurrent readpage would fail to serialize with the bh
613 * and it would read from disk before we reach the platter.
614 */
615 if (buffer_heads_over_limit && PageUptodate(page))
616 try_to_free_buffers(page);
617 }
618
619 BUG_ON(PageWriteback(page));
620 set_page_writeback(page);
621 unlock_page(page);
622 if (boundary || (first_unmapped != blocks_per_page)) {
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 = blocks[blocks_per_page - 1];
630 }
631 goto out;
632
633confused:
634 if (bio)
635 bio = mpage_bio_submit(WRITE, bio);
636
637 if (mpd->use_writepage) {
638 ret = mapping->a_ops->writepage(page, wbc);
639 } else {
640 ret = -EAGAIN;
641 goto out;
642 }
643 /*
644 * The caller has a ref on the inode, so *mapping is stable
645 */
646 mapping_set_error(mapping, ret);
647out:
648 mpd->bio = bio;
649 return ret;
650}
651
652/**
653 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
654 * @mapping: address space structure to write
655 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
656 * @get_block: the filesystem's block mapper function.
657 * If this is NULL then use a_ops->writepage. Otherwise, go
658 * direct-to-BIO.
659 *
660 * This is a library function, which implements the writepages()
661 * address_space_operation.
662 *
663 * If a page is already under I/O, generic_writepages() skips it, even
664 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
665 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
666 * and msync() need to guarantee that all the data which was dirty at the time
667 * the call was made get new I/O started against them. If wbc->sync_mode is
668 * WB_SYNC_ALL then we were called for data integrity and we must wait for
669 * existing IO to complete.
670 */
671int
672mpage_writepages(struct address_space *mapping,
673 struct writeback_control *wbc, get_block_t get_block)
674{
675 struct blk_plug plug;
676 int ret;
677
678 blk_start_plug(&plug);
679
680 if (!get_block)
681 ret = generic_writepages(mapping, wbc);
682 else {
683 struct mpage_data mpd = {
684 .bio = NULL,
685 .last_block_in_bio = 0,
686 .get_block = get_block,
687 .use_writepage = 1,
688 };
689
690 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
691 if (mpd.bio)
692 mpage_bio_submit(WRITE, mpd.bio);
693 }
694 blk_finish_plug(&plug);
695 return ret;
696}
697EXPORT_SYMBOL(mpage_writepages);
698
699int mpage_writepage(struct page *page, get_block_t get_block,
700 struct writeback_control *wbc)
701{
702 struct mpage_data mpd = {
703 .bio = NULL,
704 .last_block_in_bio = 0,
705 .get_block = get_block,
706 .use_writepage = 0,
707 };
708 int ret = __mpage_writepage(page, wbc, &mpd);
709 if (mpd.bio)
710 mpage_bio_submit(WRITE, mpd.bio);
711 return ret;
712}
713EXPORT_SYMBOL(mpage_writepage);