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_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);