4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains functions related to preparing and submitting BIOs which contain
7 * multiple pagecache pages.
9 * 15May2002 Andrew Morton
11 * 27Jun2002 axboe@suse.de
12 * use bio_add_page() to build bio's just the right size
15 #include <linux/kernel.h>
16 #include <linux/export.h>
18 #include <linux/kdev_t.h>
19 #include <linux/gfp.h>
20 #include <linux/bio.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>
33 #define CREATE_TRACE_POINTS
34 #include <trace/events/android_fs.h>
36 EXPORT_TRACEPOINT_SYMBOL(android_fs_datawrite_start);
37 EXPORT_TRACEPOINT_SYMBOL(android_fs_datawrite_end);
38 EXPORT_TRACEPOINT_SYMBOL(android_fs_dataread_start);
39 EXPORT_TRACEPOINT_SYMBOL(android_fs_dataread_end);
42 * I/O completion handler for multipage BIOs.
44 * The mpage code never puts partial pages into a BIO (except for end-of-file).
45 * If a page does not map to a contiguous run of blocks then it simply falls
46 * back to block_read_full_page().
48 * Why is this? If a page's completion depends on a number of different BIOs
49 * which can complete in any order (or at the same time) then determining the
50 * status of that page is hard. See end_buffer_async_read() for the details.
51 * There is no point in duplicating all that complexity.
53 static void mpage_end_io(struct bio *bio)
58 if (trace_android_fs_dataread_end_enabled() &&
59 (bio_data_dir(bio) == READ)) {
60 struct page *first_page = bio->bi_io_vec[0].bv_page;
62 if (first_page != NULL)
63 trace_android_fs_dataread_end(first_page->mapping->host,
64 page_offset(first_page),
65 bio->bi_iter.bi_size);
68 bio_for_each_segment_all(bv, bio, i) {
69 struct page *page = bv->bv_page;
70 page_endio(page, bio_data_dir(bio), bio->bi_error);
76 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
78 if (trace_android_fs_dataread_start_enabled() && (rw == READ)) {
79 struct page *first_page = bio->bi_io_vec[0].bv_page;
81 if (first_page != NULL) {
82 char *path, pathbuf[MAX_TRACE_PATHBUF_LEN];
84 path = android_fstrace_get_pathname(pathbuf,
85 MAX_TRACE_PATHBUF_LEN,
86 first_page->mapping->host);
87 trace_android_fs_dataread_start(
88 first_page->mapping->host,
89 page_offset(first_page),
96 bio->bi_end_io = mpage_end_io;
97 guard_bio_eod(rw, bio);
103 mpage_alloc(struct block_device *bdev,
104 sector_t first_sector, int nr_vecs,
109 bio = bio_alloc(gfp_flags, nr_vecs);
111 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
112 while (!bio && (nr_vecs /= 2))
113 bio = bio_alloc(gfp_flags, nr_vecs);
118 bio->bi_iter.bi_sector = first_sector;
124 * support function for mpage_readpages. The fs supplied get_block might
125 * return an up to date buffer. This is used to map that buffer into
126 * the page, which allows readpage to avoid triggering a duplicate call
129 * The idea is to avoid adding buffers to pages that don't already have
130 * them. So when the buffer is up to date and the page size == block size,
131 * this marks the page up to date instead of adding new buffers.
134 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
136 struct inode *inode = page->mapping->host;
137 struct buffer_head *page_bh, *head;
140 if (!page_has_buffers(page)) {
142 * don't make any buffers if there is only one buffer on
143 * the page and the page just needs to be set up to date
145 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
146 buffer_uptodate(bh)) {
147 SetPageUptodate(page);
150 create_empty_buffers(page, i_blocksize(inode), 0);
152 head = page_buffers(page);
155 if (block == page_block) {
156 page_bh->b_state = bh->b_state;
157 page_bh->b_bdev = bh->b_bdev;
158 page_bh->b_blocknr = bh->b_blocknr;
161 page_bh = page_bh->b_this_page;
163 } while (page_bh != head);
167 * This is the worker routine which does all the work of mapping the disk
168 * blocks and constructs largest possible bios, submits them for IO if the
169 * blocks are not contiguous on the disk.
171 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
172 * represent the validity of its disk mapping and to decide when to do the next
176 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
177 sector_t *last_block_in_bio, struct buffer_head *map_bh,
178 unsigned long *first_logical_block, get_block_t get_block,
181 struct inode *inode = page->mapping->host;
182 const unsigned blkbits = inode->i_blkbits;
183 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
184 const unsigned blocksize = 1 << blkbits;
185 sector_t block_in_file;
187 sector_t last_block_in_file;
188 sector_t blocks[MAX_BUF_PER_PAGE];
190 unsigned first_hole = blocks_per_page;
191 struct block_device *bdev = NULL;
193 int fully_mapped = 1;
195 unsigned relative_block;
197 if (page_has_buffers(page))
200 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
201 last_block = block_in_file + nr_pages * blocks_per_page;
202 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
203 if (last_block > last_block_in_file)
204 last_block = last_block_in_file;
208 * Map blocks using the result from the previous get_blocks call first.
210 nblocks = map_bh->b_size >> blkbits;
211 if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
212 block_in_file < (*first_logical_block + nblocks)) {
213 unsigned map_offset = block_in_file - *first_logical_block;
214 unsigned last = nblocks - map_offset;
216 for (relative_block = 0; ; relative_block++) {
217 if (relative_block == last) {
218 clear_buffer_mapped(map_bh);
221 if (page_block == blocks_per_page)
223 blocks[page_block] = map_bh->b_blocknr + map_offset +
228 bdev = map_bh->b_bdev;
232 * Then do more get_blocks calls until we are done with this page.
234 map_bh->b_page = page;
235 while (page_block < blocks_per_page) {
239 if (block_in_file < last_block) {
240 map_bh->b_size = (last_block-block_in_file) << blkbits;
241 if (get_block(inode, block_in_file, map_bh, 0))
243 *first_logical_block = block_in_file;
246 if (!buffer_mapped(map_bh)) {
248 if (first_hole == blocks_per_page)
249 first_hole = page_block;
255 /* some filesystems will copy data into the page during
256 * the get_block call, in which case we don't want to
257 * read it again. map_buffer_to_page copies the data
258 * we just collected from get_block into the page's buffers
259 * so readpage doesn't have to repeat the get_block call
261 if (buffer_uptodate(map_bh)) {
262 map_buffer_to_page(page, map_bh, page_block);
266 if (first_hole != blocks_per_page)
267 goto confused; /* hole -> non-hole */
269 /* Contiguous blocks? */
270 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
272 nblocks = map_bh->b_size >> blkbits;
273 for (relative_block = 0; ; relative_block++) {
274 if (relative_block == nblocks) {
275 clear_buffer_mapped(map_bh);
277 } else if (page_block == blocks_per_page)
279 blocks[page_block] = map_bh->b_blocknr+relative_block;
283 bdev = map_bh->b_bdev;
286 if (first_hole != blocks_per_page) {
287 zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
288 if (first_hole == 0) {
289 SetPageUptodate(page);
293 } else if (fully_mapped) {
294 SetPageMappedToDisk(page);
297 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
298 cleancache_get_page(page) == 0) {
299 SetPageUptodate(page);
304 * This page will go to BIO. Do we need to send this BIO off first?
306 if (bio && (*last_block_in_bio != blocks[0] - 1))
307 bio = mpage_bio_submit(READ, bio);
311 if (first_hole == blocks_per_page) {
312 if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
316 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
317 min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
322 length = first_hole << blkbits;
323 if (bio_add_page(bio, page, length, 0) < length) {
324 bio = mpage_bio_submit(READ, bio);
328 relative_block = block_in_file - *first_logical_block;
329 nblocks = map_bh->b_size >> blkbits;
330 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
331 (first_hole != blocks_per_page))
332 bio = mpage_bio_submit(READ, bio);
334 *last_block_in_bio = blocks[blocks_per_page - 1];
340 bio = mpage_bio_submit(READ, bio);
341 if (!PageUptodate(page))
342 block_read_full_page(page, get_block);
349 * mpage_readpages - populate an address space with some pages & start reads against them
350 * @mapping: the address_space
351 * @pages: The address of a list_head which contains the target pages. These
352 * pages have their ->index populated and are otherwise uninitialised.
353 * The page at @pages->prev has the lowest file offset, and reads should be
354 * issued in @pages->prev to @pages->next order.
355 * @nr_pages: The number of pages at *@pages
356 * @get_block: The filesystem's block mapper function.
358 * This function walks the pages and the blocks within each page, building and
359 * emitting large BIOs.
361 * If anything unusual happens, such as:
363 * - encountering a page which has buffers
364 * - encountering a page which has a non-hole after a hole
365 * - encountering a page with non-contiguous blocks
367 * then this code just gives up and calls the buffer_head-based read function.
368 * It does handle a page which has holes at the end - that is a common case:
369 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
371 * BH_Boundary explanation:
373 * There is a problem. The mpage read code assembles several pages, gets all
374 * their disk mappings, and then submits them all. That's fine, but obtaining
375 * the disk mappings may require I/O. Reads of indirect blocks, for example.
377 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
378 * submitted in the following order:
379 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
381 * because the indirect block has to be read to get the mappings of blocks
382 * 13,14,15,16. Obviously, this impacts performance.
384 * So what we do it to allow the filesystem's get_block() function to set
385 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
386 * after this one will require I/O against a block which is probably close to
387 * this one. So you should push what I/O you have currently accumulated.
389 * This all causes the disk requests to be issued in the correct order.
392 mpage_readpages(struct address_space *mapping, struct list_head *pages,
393 unsigned nr_pages, get_block_t get_block)
395 struct bio *bio = NULL;
397 sector_t last_block_in_bio = 0;
398 struct buffer_head map_bh;
399 unsigned long first_logical_block = 0;
400 gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
404 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
405 struct page *page = list_entry(pages->prev, struct page, lru);
407 prefetchw(&page->flags);
408 list_del(&page->lru);
409 if (!add_to_page_cache_lru(page, mapping,
412 bio = do_mpage_readpage(bio, page,
414 &last_block_in_bio, &map_bh,
415 &first_logical_block,
418 page_cache_release(page);
420 BUG_ON(!list_empty(pages));
422 mpage_bio_submit(READ, bio);
425 EXPORT_SYMBOL(mpage_readpages);
428 * This isn't called much at all
430 int mpage_readpage(struct page *page, get_block_t get_block)
432 struct bio *bio = NULL;
433 sector_t last_block_in_bio = 0;
434 struct buffer_head map_bh;
435 unsigned long first_logical_block = 0;
436 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
440 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
441 &map_bh, &first_logical_block, get_block, gfp);
443 mpage_bio_submit(READ, bio);
446 EXPORT_SYMBOL(mpage_readpage);
449 * Writing is not so simple.
451 * If the page has buffers then they will be used for obtaining the disk
452 * mapping. We only support pages which are fully mapped-and-dirty, with a
453 * special case for pages which are unmapped at the end: end-of-file.
455 * If the page has no buffers (preferred) then the page is mapped here.
457 * If all blocks are found to be contiguous then the page can go into the
458 * BIO. Otherwise fall back to the mapping's writepage().
460 * FIXME: This code wants an estimate of how many pages are still to be
461 * written, so it can intelligently allocate a suitably-sized BIO. For now,
462 * just allocate full-size (16-page) BIOs.
467 sector_t last_block_in_bio;
468 get_block_t *get_block;
469 unsigned use_writepage;
473 * We have our BIO, so we can now mark the buffers clean. Make
474 * sure to only clean buffers which we know we'll be writing.
476 static void clean_buffers(struct page *page, unsigned first_unmapped)
478 unsigned buffer_counter = 0;
479 struct buffer_head *bh, *head;
480 if (!page_has_buffers(page))
482 head = page_buffers(page);
486 if (buffer_counter++ == first_unmapped)
488 clear_buffer_dirty(bh);
489 bh = bh->b_this_page;
490 } while (bh != head);
493 * we cannot drop the bh if the page is not uptodate or a concurrent
494 * readpage would fail to serialize with the bh and it would read from
495 * disk before we reach the platter.
497 if (buffer_heads_over_limit && PageUptodate(page))
498 try_to_free_buffers(page);
501 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
504 struct mpage_data *mpd = data;
505 struct bio *bio = mpd->bio;
506 struct address_space *mapping = page->mapping;
507 struct inode *inode = page->mapping->host;
508 const unsigned blkbits = inode->i_blkbits;
509 unsigned long end_index;
510 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
512 sector_t block_in_file;
513 sector_t blocks[MAX_BUF_PER_PAGE];
515 unsigned first_unmapped = blocks_per_page;
516 struct block_device *bdev = NULL;
518 sector_t boundary_block = 0;
519 struct block_device *boundary_bdev = NULL;
521 struct buffer_head map_bh;
522 loff_t i_size = i_size_read(inode);
524 int wr = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
526 if (page_has_buffers(page)) {
527 struct buffer_head *head = page_buffers(page);
528 struct buffer_head *bh = head;
530 /* If they're all mapped and dirty, do it */
533 BUG_ON(buffer_locked(bh));
534 if (!buffer_mapped(bh)) {
536 * unmapped dirty buffers are created by
537 * __set_page_dirty_buffers -> mmapped data
539 if (buffer_dirty(bh))
541 if (first_unmapped == blocks_per_page)
542 first_unmapped = page_block;
546 if (first_unmapped != blocks_per_page)
547 goto confused; /* hole -> non-hole */
549 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
552 if (bh->b_blocknr != blocks[page_block-1] + 1)
555 blocks[page_block++] = bh->b_blocknr;
556 boundary = buffer_boundary(bh);
558 boundary_block = bh->b_blocknr;
559 boundary_bdev = bh->b_bdev;
562 } while ((bh = bh->b_this_page) != head);
568 * Page has buffers, but they are all unmapped. The page was
569 * created by pagein or read over a hole which was handled by
570 * block_read_full_page(). If this address_space is also
571 * using mpage_readpages then this can rarely happen.
577 * The page has no buffers: map it to disk
579 BUG_ON(!PageUptodate(page));
580 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
581 last_block = (i_size - 1) >> blkbits;
582 map_bh.b_page = page;
583 for (page_block = 0; page_block < blocks_per_page; ) {
586 map_bh.b_size = 1 << blkbits;
587 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
589 if (buffer_new(&map_bh))
590 unmap_underlying_metadata(map_bh.b_bdev,
592 if (buffer_boundary(&map_bh)) {
593 boundary_block = map_bh.b_blocknr;
594 boundary_bdev = map_bh.b_bdev;
597 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
600 blocks[page_block++] = map_bh.b_blocknr;
601 boundary = buffer_boundary(&map_bh);
602 bdev = map_bh.b_bdev;
603 if (block_in_file == last_block)
607 BUG_ON(page_block == 0);
609 first_unmapped = page_block;
612 end_index = i_size >> PAGE_CACHE_SHIFT;
613 if (page->index >= end_index) {
615 * The page straddles i_size. It must be zeroed out on each
616 * and every writepage invocation because it may be mmapped.
617 * "A file is mapped in multiples of the page size. For a file
618 * that is not a multiple of the page size, the remaining memory
619 * is zeroed when mapped, and writes to that region are not
620 * written out to the file."
622 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
624 if (page->index > end_index || !offset)
626 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
630 * This page will go to BIO. Do we need to send this BIO off first?
632 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
633 bio = mpage_bio_submit(wr, bio);
637 if (first_unmapped == blocks_per_page) {
638 if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
640 clean_buffers(page, first_unmapped);
644 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
645 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
649 wbc_init_bio(wbc, bio);
653 * Must try to add the page before marking the buffer clean or
654 * the confused fail path above (OOM) will be very confused when
655 * it finds all bh marked clean (i.e. it will not write anything)
657 wbc_account_io(wbc, page, PAGE_SIZE);
658 length = first_unmapped << blkbits;
659 if (bio_add_page(bio, page, length, 0) < length) {
660 bio = mpage_bio_submit(wr, bio);
664 clean_buffers(page, first_unmapped);
666 BUG_ON(PageWriteback(page));
667 set_page_writeback(page);
669 if (boundary || (first_unmapped != blocks_per_page)) {
670 bio = mpage_bio_submit(wr, bio);
671 if (boundary_block) {
672 write_boundary_block(boundary_bdev,
673 boundary_block, 1 << blkbits);
676 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
682 bio = mpage_bio_submit(wr, bio);
684 if (mpd->use_writepage) {
685 ret = mapping->a_ops->writepage(page, wbc);
691 * The caller has a ref on the inode, so *mapping is stable
693 mapping_set_error(mapping, ret);
700 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
701 * @mapping: address space structure to write
702 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
703 * @get_block: the filesystem's block mapper function.
704 * If this is NULL then use a_ops->writepage. Otherwise, go
707 * This is a library function, which implements the writepages()
708 * address_space_operation.
710 * If a page is already under I/O, generic_writepages() skips it, even
711 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
712 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
713 * and msync() need to guarantee that all the data which was dirty at the time
714 * the call was made get new I/O started against them. If wbc->sync_mode is
715 * WB_SYNC_ALL then we were called for data integrity and we must wait for
716 * existing IO to complete.
719 mpage_writepages(struct address_space *mapping,
720 struct writeback_control *wbc, get_block_t get_block)
722 struct blk_plug plug;
725 blk_start_plug(&plug);
728 ret = generic_writepages(mapping, wbc);
730 struct mpage_data mpd = {
732 .last_block_in_bio = 0,
733 .get_block = get_block,
737 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
739 int wr = (wbc->sync_mode == WB_SYNC_ALL ?
741 mpage_bio_submit(wr, mpd.bio);
744 blk_finish_plug(&plug);
747 EXPORT_SYMBOL(mpage_writepages);
749 int mpage_writepage(struct page *page, get_block_t get_block,
750 struct writeback_control *wbc)
752 struct mpage_data mpd = {
754 .last_block_in_bio = 0,
755 .get_block = get_block,
758 int ret = __mpage_writepage(page, wbc, &mpd);
760 int wr = (wbc->sync_mode == WB_SYNC_ALL ?
762 mpage_bio_submit(wr, mpd.bio);
766 EXPORT_SYMBOL(mpage_writepage);