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