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 * I/O completion handler for multipage BIOs.
35 * The mpage code never puts partial pages into a BIO (except for end-of-file).
36 * If a page does not map to a contiguous run of blocks then it simply falls
37 * back to block_read_full_page().
39 * Why is this? If a page's completion depends on a number of different BIOs
40 * which can complete in any order (or at the same time) then determining the
41 * status of that page is hard. See end_buffer_async_read() for the details.
42 * There is no point in duplicating all that complexity.
44 static void mpage_end_io(struct bio *bio, int err)
46 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
47 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
50 struct page *page = bvec->bv_page;
52 if (--bvec >= bio->bi_io_vec)
53 prefetchw(&bvec->bv_page->flags);
54 if (bio_data_dir(bio) == READ) {
56 SetPageUptodate(page);
58 ClearPageUptodate(page);
62 } else { /* bio_data_dir(bio) == WRITE */
66 set_bit(AS_EIO, &page->mapping->flags);
68 end_page_writeback(page);
70 } while (bvec >= bio->bi_io_vec);
74 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
76 bio->bi_end_io = mpage_end_io;
82 mpage_alloc(struct block_device *bdev,
83 sector_t first_sector, int nr_vecs,
88 bio = bio_alloc(gfp_flags, nr_vecs);
90 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
91 while (!bio && (nr_vecs /= 2))
92 bio = bio_alloc(gfp_flags, nr_vecs);
97 bio->bi_sector = first_sector;
103 * support function for mpage_readpages. The fs supplied get_block might
104 * return an up to date buffer. This is used to map that buffer into
105 * the page, which allows readpage to avoid triggering a duplicate call
108 * The idea is to avoid adding buffers to pages that don't already have
109 * them. So when the buffer is up to date and the page size == block size,
110 * this marks the page up to date instead of adding new buffers.
113 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
115 struct inode *inode = page->mapping->host;
116 struct buffer_head *page_bh, *head;
119 if (!page_has_buffers(page)) {
121 * don't make any buffers if there is only one buffer on
122 * the page and the page just needs to be set up to date
124 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
125 buffer_uptodate(bh)) {
126 SetPageUptodate(page);
129 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
131 head = page_buffers(page);
134 if (block == page_block) {
135 page_bh->b_state = bh->b_state;
136 page_bh->b_bdev = bh->b_bdev;
137 page_bh->b_blocknr = bh->b_blocknr;
140 page_bh = page_bh->b_this_page;
142 } while (page_bh != head);
146 * This is the worker routine which does all the work of mapping the disk
147 * blocks and constructs largest possible bios, submits them for IO if the
148 * blocks are not contiguous on the disk.
150 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
151 * represent the validity of its disk mapping and to decide when to do the next
155 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
156 sector_t *last_block_in_bio, struct buffer_head *map_bh,
157 unsigned long *first_logical_block, get_block_t get_block)
159 struct inode *inode = page->mapping->host;
160 const unsigned blkbits = inode->i_blkbits;
161 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
162 const unsigned blocksize = 1 << blkbits;
163 sector_t block_in_file;
165 sector_t last_block_in_file;
166 sector_t blocks[MAX_BUF_PER_PAGE];
168 unsigned first_hole = blocks_per_page;
169 struct block_device *bdev = NULL;
171 int fully_mapped = 1;
173 unsigned relative_block;
175 if (page_has_buffers(page))
178 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
179 last_block = block_in_file + nr_pages * blocks_per_page;
180 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
181 if (last_block > last_block_in_file)
182 last_block = last_block_in_file;
186 * Map blocks using the result from the previous get_blocks call first.
188 nblocks = map_bh->b_size >> blkbits;
189 if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
190 block_in_file < (*first_logical_block + nblocks)) {
191 unsigned map_offset = block_in_file - *first_logical_block;
192 unsigned last = nblocks - map_offset;
194 for (relative_block = 0; ; relative_block++) {
195 if (relative_block == last) {
196 clear_buffer_mapped(map_bh);
199 if (page_block == blocks_per_page)
201 blocks[page_block] = map_bh->b_blocknr + map_offset +
206 bdev = map_bh->b_bdev;
210 * Then do more get_blocks calls until we are done with this page.
212 map_bh->b_page = page;
213 while (page_block < blocks_per_page) {
217 if (block_in_file < last_block) {
218 map_bh->b_size = (last_block-block_in_file) << blkbits;
219 if (get_block(inode, block_in_file, map_bh, 0))
221 *first_logical_block = block_in_file;
224 if (!buffer_mapped(map_bh)) {
226 if (first_hole == blocks_per_page)
227 first_hole = page_block;
233 /* some filesystems will copy data into the page during
234 * the get_block call, in which case we don't want to
235 * read it again. map_buffer_to_page copies the data
236 * we just collected from get_block into the page's buffers
237 * so readpage doesn't have to repeat the get_block call
239 if (buffer_uptodate(map_bh)) {
240 map_buffer_to_page(page, map_bh, page_block);
244 if (first_hole != blocks_per_page)
245 goto confused; /* hole -> non-hole */
247 /* Contiguous blocks? */
248 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
250 nblocks = map_bh->b_size >> blkbits;
251 for (relative_block = 0; ; relative_block++) {
252 if (relative_block == nblocks) {
253 clear_buffer_mapped(map_bh);
255 } else if (page_block == blocks_per_page)
257 blocks[page_block] = map_bh->b_blocknr+relative_block;
261 bdev = map_bh->b_bdev;
264 if (first_hole != blocks_per_page) {
265 zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
266 if (first_hole == 0) {
267 SetPageUptodate(page);
271 } else if (fully_mapped) {
272 SetPageMappedToDisk(page);
275 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
276 cleancache_get_page(page) == 0) {
277 SetPageUptodate(page);
282 * This page will go to BIO. Do we need to send this BIO off first?
284 if (bio && (*last_block_in_bio != blocks[0] - 1))
285 bio = mpage_bio_submit(READ, bio);
289 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
290 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
296 length = first_hole << blkbits;
297 if (bio_add_page(bio, page, length, 0) < length) {
298 bio = mpage_bio_submit(READ, bio);
302 relative_block = block_in_file - *first_logical_block;
303 nblocks = map_bh->b_size >> blkbits;
304 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
305 (first_hole != blocks_per_page))
306 bio = mpage_bio_submit(READ, bio);
308 *last_block_in_bio = blocks[blocks_per_page - 1];
314 bio = mpage_bio_submit(READ, bio);
315 if (!PageUptodate(page))
316 block_read_full_page(page, get_block);
323 * mpage_readpages - populate an address space with some pages & start reads against them
324 * @mapping: the address_space
325 * @pages: The address of a list_head which contains the target pages. These
326 * pages have their ->index populated and are otherwise uninitialised.
327 * The page at @pages->prev has the lowest file offset, and reads should be
328 * issued in @pages->prev to @pages->next order.
329 * @nr_pages: The number of pages at *@pages
330 * @get_block: The filesystem's block mapper function.
332 * This function walks the pages and the blocks within each page, building and
333 * emitting large BIOs.
335 * If anything unusual happens, such as:
337 * - encountering a page which has buffers
338 * - encountering a page which has a non-hole after a hole
339 * - encountering a page with non-contiguous blocks
341 * then this code just gives up and calls the buffer_head-based read function.
342 * It does handle a page which has holes at the end - that is a common case:
343 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
345 * BH_Boundary explanation:
347 * There is a problem. The mpage read code assembles several pages, gets all
348 * their disk mappings, and then submits them all. That's fine, but obtaining
349 * the disk mappings may require I/O. Reads of indirect blocks, for example.
351 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
352 * submitted in the following order:
353 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
355 * because the indirect block has to be read to get the mappings of blocks
356 * 13,14,15,16. Obviously, this impacts performance.
358 * So what we do it to allow the filesystem's get_block() function to set
359 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
360 * after this one will require I/O against a block which is probably close to
361 * this one. So you should push what I/O you have currently accumulated.
363 * This all causes the disk requests to be issued in the correct order.
366 mpage_readpages(struct address_space *mapping, struct list_head *pages,
367 unsigned nr_pages, get_block_t get_block)
369 struct bio *bio = NULL;
371 sector_t last_block_in_bio = 0;
372 struct buffer_head map_bh;
373 unsigned long first_logical_block = 0;
377 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
378 struct page *page = list_entry(pages->prev, struct page, lru);
380 prefetchw(&page->flags);
381 list_del(&page->lru);
382 if (!add_to_page_cache_lru(page, mapping,
383 page->index, GFP_KERNEL)) {
384 bio = do_mpage_readpage(bio, page,
386 &last_block_in_bio, &map_bh,
387 &first_logical_block,
390 page_cache_release(page);
392 BUG_ON(!list_empty(pages));
394 mpage_bio_submit(READ, bio);
397 EXPORT_SYMBOL(mpage_readpages);
400 * This isn't called much at all
402 int mpage_readpage(struct page *page, get_block_t get_block)
404 struct bio *bio = NULL;
405 sector_t last_block_in_bio = 0;
406 struct buffer_head map_bh;
407 unsigned long first_logical_block = 0;
411 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
412 &map_bh, &first_logical_block, get_block);
414 mpage_bio_submit(READ, bio);
417 EXPORT_SYMBOL(mpage_readpage);
420 * Writing is not so simple.
422 * If the page has buffers then they will be used for obtaining the disk
423 * mapping. We only support pages which are fully mapped-and-dirty, with a
424 * special case for pages which are unmapped at the end: end-of-file.
426 * If the page has no buffers (preferred) then the page is mapped here.
428 * If all blocks are found to be contiguous then the page can go into the
429 * BIO. Otherwise fall back to the mapping's writepage().
431 * FIXME: This code wants an estimate of how many pages are still to be
432 * written, so it can intelligently allocate a suitably-sized BIO. For now,
433 * just allocate full-size (16-page) BIOs.
438 sector_t last_block_in_bio;
439 get_block_t *get_block;
440 unsigned use_writepage;
443 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
446 struct mpage_data *mpd = data;
447 struct bio *bio = mpd->bio;
448 struct address_space *mapping = page->mapping;
449 struct inode *inode = page->mapping->host;
450 const unsigned blkbits = inode->i_blkbits;
451 unsigned long end_index;
452 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
454 sector_t block_in_file;
455 sector_t blocks[MAX_BUF_PER_PAGE];
457 unsigned first_unmapped = blocks_per_page;
458 struct block_device *bdev = NULL;
460 sector_t boundary_block = 0;
461 struct block_device *boundary_bdev = NULL;
463 struct buffer_head map_bh;
464 loff_t i_size = i_size_read(inode);
467 if (page_has_buffers(page)) {
468 struct buffer_head *head = page_buffers(page);
469 struct buffer_head *bh = head;
471 /* If they're all mapped and dirty, do it */
474 BUG_ON(buffer_locked(bh));
475 if (!buffer_mapped(bh)) {
477 * unmapped dirty buffers are created by
478 * __set_page_dirty_buffers -> mmapped data
480 if (buffer_dirty(bh))
482 if (first_unmapped == blocks_per_page)
483 first_unmapped = page_block;
487 if (first_unmapped != blocks_per_page)
488 goto confused; /* hole -> non-hole */
490 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
493 if (bh->b_blocknr != blocks[page_block-1] + 1)
496 blocks[page_block++] = bh->b_blocknr;
497 boundary = buffer_boundary(bh);
499 boundary_block = bh->b_blocknr;
500 boundary_bdev = bh->b_bdev;
503 } while ((bh = bh->b_this_page) != head);
509 * Page has buffers, but they are all unmapped. The page was
510 * created by pagein or read over a hole which was handled by
511 * block_read_full_page(). If this address_space is also
512 * using mpage_readpages then this can rarely happen.
518 * The page has no buffers: map it to disk
520 BUG_ON(!PageUptodate(page));
521 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
522 last_block = (i_size - 1) >> blkbits;
523 map_bh.b_page = page;
524 for (page_block = 0; page_block < blocks_per_page; ) {
527 map_bh.b_size = 1 << blkbits;
528 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
530 if (buffer_new(&map_bh))
531 unmap_underlying_metadata(map_bh.b_bdev,
533 if (buffer_boundary(&map_bh)) {
534 boundary_block = map_bh.b_blocknr;
535 boundary_bdev = map_bh.b_bdev;
538 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
541 blocks[page_block++] = map_bh.b_blocknr;
542 boundary = buffer_boundary(&map_bh);
543 bdev = map_bh.b_bdev;
544 if (block_in_file == last_block)
548 BUG_ON(page_block == 0);
550 first_unmapped = page_block;
553 end_index = i_size >> PAGE_CACHE_SHIFT;
554 if (page->index >= end_index) {
556 * The page straddles i_size. It must be zeroed out on each
557 * and every writepage invocation because it may be mmapped.
558 * "A file is mapped in multiples of the page size. For a file
559 * that is not a multiple of the page size, the remaining memory
560 * is zeroed when mapped, and writes to that region are not
561 * written out to the file."
563 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
565 if (page->index > end_index || !offset)
567 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
571 * This page will go to BIO. Do we need to send this BIO off first?
573 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
574 bio = mpage_bio_submit(WRITE, bio);
578 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
579 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
585 * Must try to add the page before marking the buffer clean or
586 * the confused fail path above (OOM) will be very confused when
587 * it finds all bh marked clean (i.e. it will not write anything)
589 length = first_unmapped << blkbits;
590 if (bio_add_page(bio, page, length, 0) < length) {
591 bio = mpage_bio_submit(WRITE, bio);
596 * OK, we have our BIO, so we can now mark the buffers clean. Make
597 * sure to only clean buffers which we know we'll be writing.
599 if (page_has_buffers(page)) {
600 struct buffer_head *head = page_buffers(page);
601 struct buffer_head *bh = head;
602 unsigned buffer_counter = 0;
605 if (buffer_counter++ == first_unmapped)
607 clear_buffer_dirty(bh);
608 bh = bh->b_this_page;
609 } while (bh != head);
612 * we cannot drop the bh if the page is not uptodate
613 * or a concurrent readpage would fail to serialize with the bh
614 * and it would read from disk before we reach the platter.
616 if (buffer_heads_over_limit && PageUptodate(page))
617 try_to_free_buffers(page);
620 BUG_ON(PageWriteback(page));
621 set_page_writeback(page);
623 if (boundary || (first_unmapped != blocks_per_page)) {
624 bio = mpage_bio_submit(WRITE, bio);
625 if (boundary_block) {
626 write_boundary_block(boundary_bdev,
627 boundary_block, 1 << blkbits);
630 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
636 bio = mpage_bio_submit(WRITE, bio);
638 if (mpd->use_writepage) {
639 ret = mapping->a_ops->writepage(page, wbc);
645 * The caller has a ref on the inode, so *mapping is stable
647 mapping_set_error(mapping, ret);
654 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
655 * @mapping: address space structure to write
656 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
657 * @get_block: the filesystem's block mapper function.
658 * If this is NULL then use a_ops->writepage. Otherwise, go
661 * This is a library function, which implements the writepages()
662 * address_space_operation.
664 * If a page is already under I/O, generic_writepages() skips it, even
665 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
666 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
667 * and msync() need to guarantee that all the data which was dirty at the time
668 * the call was made get new I/O started against them. If wbc->sync_mode is
669 * WB_SYNC_ALL then we were called for data integrity and we must wait for
670 * existing IO to complete.
673 mpage_writepages(struct address_space *mapping,
674 struct writeback_control *wbc, get_block_t get_block)
676 struct blk_plug plug;
679 blk_start_plug(&plug);
682 ret = generic_writepages(mapping, wbc);
684 struct mpage_data mpd = {
686 .last_block_in_bio = 0,
687 .get_block = get_block,
691 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
693 mpage_bio_submit(WRITE, mpd.bio);
695 blk_finish_plug(&plug);
698 EXPORT_SYMBOL(mpage_writepages);
700 int mpage_writepage(struct page *page, get_block_t get_block,
701 struct writeback_control *wbc)
703 struct mpage_data mpd = {
705 .last_block_in_bio = 0,
706 .get_block = get_block,
709 int ret = __mpage_writepage(page, wbc, &mpd);
711 mpage_bio_submit(WRITE, mpd.bio);
714 EXPORT_SYMBOL(mpage_writepage);