4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains functions related to preparing and submitting BIOs which contain
7 * multiple pagecache pages.
9 * 15May2002 akpm@zip.com.au
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/module.h>
18 #include <linux/kdev_t.h>
19 #include <linux/bio.h>
21 #include <linux/buffer_head.h>
22 #include <linux/blkdev.h>
23 #include <linux/highmem.h>
24 #include <linux/prefetch.h>
25 #include <linux/mpage.h>
26 #include <linux/writeback.h>
27 #include <linux/backing-dev.h>
28 #include <linux/pagevec.h>
31 * I/O completion handler for multipage BIOs.
33 * The mpage code never puts partial pages into a BIO (except for end-of-file).
34 * If a page does not map to a contiguous run of blocks then it simply falls
35 * back to block_read_full_page().
37 * Why is this? If a page's completion depends on a number of different BIOs
38 * which can complete in any order (or at the same time) then determining the
39 * status of that page is hard. See end_buffer_async_read() for the details.
40 * There is no point in duplicating all that complexity.
42 static int mpage_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
44 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
45 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
51 struct page *page = bvec->bv_page;
53 if (--bvec >= bio->bi_io_vec)
54 prefetchw(&bvec->bv_page->flags);
57 SetPageUptodate(page);
59 ClearPageUptodate(page);
63 } while (bvec >= bio->bi_io_vec);
68 static int mpage_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
70 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
71 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
77 struct page *page = bvec->bv_page;
79 if (--bvec >= bio->bi_io_vec)
80 prefetchw(&bvec->bv_page->flags);
85 set_bit(AS_EIO, &page->mapping->flags);
87 end_page_writeback(page);
88 } while (bvec >= bio->bi_io_vec);
93 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
95 bio->bi_end_io = mpage_end_io_read;
97 bio->bi_end_io = mpage_end_io_write;
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_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, 1 << inode->i_blkbits, 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)
180 struct inode *inode = page->mapping->host;
181 const unsigned blkbits = inode->i_blkbits;
182 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
183 const unsigned blocksize = 1 << blkbits;
184 sector_t block_in_file;
186 sector_t last_block_in_file;
187 sector_t blocks[MAX_BUF_PER_PAGE];
189 unsigned first_hole = blocks_per_page;
190 struct block_device *bdev = NULL;
192 int fully_mapped = 1;
194 unsigned relative_block;
196 if (page_has_buffers(page))
199 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
200 last_block = block_in_file + nr_pages * blocks_per_page;
201 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
202 if (last_block > last_block_in_file)
203 last_block = last_block_in_file;
207 * Map blocks using the result from the previous get_blocks call first.
209 nblocks = map_bh->b_size >> blkbits;
210 if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
211 block_in_file < (*first_logical_block + nblocks)) {
212 unsigned map_offset = block_in_file - *first_logical_block;
213 unsigned last = nblocks - map_offset;
215 for (relative_block = 0; ; relative_block++) {
216 if (relative_block == last) {
217 clear_buffer_mapped(map_bh);
220 if (page_block == blocks_per_page)
222 blocks[page_block] = map_bh->b_blocknr + map_offset +
227 bdev = map_bh->b_bdev;
231 * Then do more get_blocks calls until we are done with this page.
233 map_bh->b_page = page;
234 while (page_block < blocks_per_page) {
238 if (block_in_file < last_block) {
239 map_bh->b_size = (last_block-block_in_file) << blkbits;
240 if (get_block(inode, block_in_file, map_bh, 0))
242 *first_logical_block = block_in_file;
245 if (!buffer_mapped(map_bh)) {
247 if (first_hole == blocks_per_page)
248 first_hole = page_block;
251 clear_buffer_mapped(map_bh);
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_page(page, first_hole << blkbits,
288 PAGE_CACHE_SIZE - (first_hole << blkbits),
290 if (first_hole == 0) {
291 SetPageUptodate(page);
295 } else if (fully_mapped) {
296 SetPageMappedToDisk(page);
300 * This page will go to BIO. Do we need to send this BIO off first?
302 if (bio && (*last_block_in_bio != blocks[0] - 1))
303 bio = mpage_bio_submit(READ, bio);
307 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
308 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
314 length = first_hole << blkbits;
315 if (bio_add_page(bio, page, length, 0) < length) {
316 bio = mpage_bio_submit(READ, bio);
320 if (buffer_boundary(map_bh) || (first_hole != blocks_per_page))
321 bio = mpage_bio_submit(READ, bio);
323 *last_block_in_bio = blocks[blocks_per_page - 1];
329 bio = mpage_bio_submit(READ, bio);
330 if (!PageUptodate(page))
331 block_read_full_page(page, get_block);
338 * mpage_readpages - populate an address space with some pages, and
339 * start reads against them.
341 * @mapping: the address_space
342 * @pages: The address of a list_head which contains the target pages. These
343 * pages have their ->index populated and are otherwise uninitialised.
345 * The page at @pages->prev has the lowest file offset, and reads should be
346 * issued in @pages->prev to @pages->next order.
348 * @nr_pages: The number of pages at *@pages
349 * @get_block: The filesystem's block mapper function.
351 * This function walks the pages and the blocks within each page, building and
352 * emitting large BIOs.
354 * If anything unusual happens, such as:
356 * - encountering a page which has buffers
357 * - encountering a page which has a non-hole after a hole
358 * - encountering a page with non-contiguous blocks
360 * then this code just gives up and calls the buffer_head-based read function.
361 * It does handle a page which has holes at the end - that is a common case:
362 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
364 * BH_Boundary explanation:
366 * There is a problem. The mpage read code assembles several pages, gets all
367 * their disk mappings, and then submits them all. That's fine, but obtaining
368 * the disk mappings may require I/O. Reads of indirect blocks, for example.
370 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
371 * submitted in the following order:
372 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
373 * because the indirect block has to be read to get the mappings of blocks
374 * 13,14,15,16. Obviously, this impacts performance.
376 * So what we do it to allow the filesystem's get_block() function to set
377 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
378 * after this one will require I/O against a block which is probably close to
379 * this one. So you should push what I/O you have currently accumulated.
381 * This all causes the disk requests to be issued in the correct order.
384 mpage_readpages(struct address_space *mapping, struct list_head *pages,
385 unsigned nr_pages, get_block_t get_block)
387 struct bio *bio = NULL;
389 sector_t last_block_in_bio = 0;
390 struct pagevec lru_pvec;
391 struct buffer_head map_bh;
392 unsigned long first_logical_block = 0;
394 clear_buffer_mapped(&map_bh);
395 pagevec_init(&lru_pvec, 0);
396 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
397 struct page *page = list_entry(pages->prev, struct page, lru);
399 prefetchw(&page->flags);
400 list_del(&page->lru);
401 if (!add_to_page_cache(page, mapping,
402 page->index, GFP_KERNEL)) {
403 bio = do_mpage_readpage(bio, page,
405 &last_block_in_bio, &map_bh,
406 &first_logical_block,
408 if (!pagevec_add(&lru_pvec, page))
409 __pagevec_lru_add(&lru_pvec);
411 page_cache_release(page);
414 pagevec_lru_add(&lru_pvec);
415 BUG_ON(!list_empty(pages));
417 mpage_bio_submit(READ, bio);
420 EXPORT_SYMBOL(mpage_readpages);
423 * This isn't called much at all
425 int mpage_readpage(struct page *page, get_block_t get_block)
427 struct bio *bio = NULL;
428 sector_t last_block_in_bio = 0;
429 struct buffer_head map_bh;
430 unsigned long first_logical_block = 0;
432 clear_buffer_mapped(&map_bh);
433 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
434 &map_bh, &first_logical_block, get_block);
436 mpage_bio_submit(READ, bio);
439 EXPORT_SYMBOL(mpage_readpage);
442 * Writing is not so simple.
444 * If the page has buffers then they will be used for obtaining the disk
445 * mapping. We only support pages which are fully mapped-and-dirty, with a
446 * special case for pages which are unmapped at the end: end-of-file.
448 * If the page has no buffers (preferred) then the page is mapped here.
450 * If all blocks are found to be contiguous then the page can go into the
451 * BIO. Otherwise fall back to the mapping's writepage().
453 * FIXME: This code wants an estimate of how many pages are still to be
454 * written, so it can intelligently allocate a suitably-sized BIO. For now,
455 * just allocate full-size (16-page) BIOs.
458 __mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
459 sector_t *last_block_in_bio, int *ret, struct writeback_control *wbc,
460 writepage_t writepage_fn)
462 struct address_space *mapping = page->mapping;
463 struct inode *inode = page->mapping->host;
464 const unsigned blkbits = inode->i_blkbits;
465 unsigned long end_index;
466 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
468 sector_t block_in_file;
469 sector_t blocks[MAX_BUF_PER_PAGE];
471 unsigned first_unmapped = blocks_per_page;
472 struct block_device *bdev = NULL;
474 sector_t boundary_block = 0;
475 struct block_device *boundary_bdev = NULL;
477 struct buffer_head map_bh;
478 loff_t i_size = i_size_read(inode);
480 if (page_has_buffers(page)) {
481 struct buffer_head *head = page_buffers(page);
482 struct buffer_head *bh = head;
484 /* If they're all mapped and dirty, do it */
487 BUG_ON(buffer_locked(bh));
488 if (!buffer_mapped(bh)) {
490 * unmapped dirty buffers are created by
491 * __set_page_dirty_buffers -> mmapped data
493 if (buffer_dirty(bh))
495 if (first_unmapped == blocks_per_page)
496 first_unmapped = page_block;
500 if (first_unmapped != blocks_per_page)
501 goto confused; /* hole -> non-hole */
503 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
506 if (bh->b_blocknr != blocks[page_block-1] + 1)
509 blocks[page_block++] = bh->b_blocknr;
510 boundary = buffer_boundary(bh);
512 boundary_block = bh->b_blocknr;
513 boundary_bdev = bh->b_bdev;
516 } while ((bh = bh->b_this_page) != head);
522 * Page has buffers, but they are all unmapped. The page was
523 * created by pagein or read over a hole which was handled by
524 * block_read_full_page(). If this address_space is also
525 * using mpage_readpages then this can rarely happen.
531 * The page has no buffers: map it to disk
533 BUG_ON(!PageUptodate(page));
534 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
535 last_block = (i_size - 1) >> blkbits;
536 map_bh.b_page = page;
537 for (page_block = 0; page_block < blocks_per_page; ) {
540 map_bh.b_size = 1 << blkbits;
541 if (get_block(inode, block_in_file, &map_bh, 1))
543 if (buffer_new(&map_bh))
544 unmap_underlying_metadata(map_bh.b_bdev,
546 if (buffer_boundary(&map_bh)) {
547 boundary_block = map_bh.b_blocknr;
548 boundary_bdev = map_bh.b_bdev;
551 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
554 blocks[page_block++] = map_bh.b_blocknr;
555 boundary = buffer_boundary(&map_bh);
556 bdev = map_bh.b_bdev;
557 if (block_in_file == last_block)
561 BUG_ON(page_block == 0);
563 first_unmapped = page_block;
566 end_index = i_size >> PAGE_CACHE_SHIFT;
567 if (page->index >= end_index) {
569 * The page straddles i_size. It must be zeroed out on each
570 * and every writepage invokation because it may be mmapped.
571 * "A file is mapped in multiples of the page size. For a file
572 * that is not a multiple of the page size, the remaining memory
573 * is zeroed when mapped, and writes to that region are not
574 * written out to the file."
576 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
578 if (page->index > end_index || !offset)
580 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset,
585 * This page will go to BIO. Do we need to send this BIO off first?
587 if (bio && *last_block_in_bio != blocks[0] - 1)
588 bio = mpage_bio_submit(WRITE, bio);
592 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
593 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
599 * Must try to add the page before marking the buffer clean or
600 * the confused fail path above (OOM) will be very confused when
601 * it finds all bh marked clean (i.e. it will not write anything)
603 length = first_unmapped << blkbits;
604 if (bio_add_page(bio, page, length, 0) < length) {
605 bio = mpage_bio_submit(WRITE, bio);
610 * OK, we have our BIO, so we can now mark the buffers clean. Make
611 * sure to only clean buffers which we know we'll be writing.
613 if (page_has_buffers(page)) {
614 struct buffer_head *head = page_buffers(page);
615 struct buffer_head *bh = head;
616 unsigned buffer_counter = 0;
619 if (buffer_counter++ == first_unmapped)
621 clear_buffer_dirty(bh);
622 bh = bh->b_this_page;
623 } while (bh != head);
626 * we cannot drop the bh if the page is not uptodate
627 * or a concurrent readpage would fail to serialize with the bh
628 * and it would read from disk before we reach the platter.
630 if (buffer_heads_over_limit && PageUptodate(page))
631 try_to_free_buffers(page);
634 BUG_ON(PageWriteback(page));
635 set_page_writeback(page);
637 if (boundary || (first_unmapped != blocks_per_page)) {
638 bio = mpage_bio_submit(WRITE, bio);
639 if (boundary_block) {
640 write_boundary_block(boundary_bdev,
641 boundary_block, 1 << blkbits);
644 *last_block_in_bio = blocks[blocks_per_page - 1];
650 bio = mpage_bio_submit(WRITE, bio);
653 *ret = (*writepage_fn)(page, wbc);
659 * The caller has a ref on the inode, so *mapping is stable
661 mapping_set_error(mapping, *ret);
667 * mpage_writepages - walk the list of dirty pages of the given
668 * address space and writepage() all of them.
670 * @mapping: address space structure to write
671 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
672 * @get_block: the filesystem's block mapper function.
673 * If this is NULL then use a_ops->writepage. Otherwise, go
676 * This is a library function, which implements the writepages()
677 * address_space_operation.
679 * If a page is already under I/O, generic_writepages() skips it, even
680 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
681 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
682 * and msync() need to guarantee that all the data which was dirty at the time
683 * the call was made get new I/O started against them. If wbc->sync_mode is
684 * WB_SYNC_ALL then we were called for data integrity and we must wait for
685 * existing IO to complete.
687 * If you fix this you should check generic_writepages() also!
690 mpage_writepages(struct address_space *mapping,
691 struct writeback_control *wbc, get_block_t get_block)
693 struct backing_dev_info *bdi = mapping->backing_dev_info;
694 struct bio *bio = NULL;
695 sector_t last_block_in_bio = 0;
698 int (*writepage)(struct page *page, struct writeback_control *wbc);
702 pgoff_t end; /* Inclusive */
706 if (wbc->nonblocking && bdi_write_congested(bdi)) {
707 wbc->encountered_congestion = 1;
712 if (get_block == NULL)
713 writepage = mapping->a_ops->writepage;
715 pagevec_init(&pvec, 0);
716 if (wbc->range_cyclic) {
717 index = mapping->writeback_index; /* Start from prev offset */
720 index = wbc->range_start >> PAGE_CACHE_SHIFT;
721 end = wbc->range_end >> PAGE_CACHE_SHIFT;
722 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
727 while (!done && (index <= end) &&
728 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
730 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
734 for (i = 0; i < nr_pages; i++) {
735 struct page *page = pvec.pages[i];
738 * At this point we hold neither mapping->tree_lock nor
739 * lock on the page itself: the page may be truncated or
740 * invalidated (changing page->mapping to NULL), or even
741 * swizzled back from swapper_space to tmpfs file
747 if (unlikely(page->mapping != mapping)) {
752 if (!wbc->range_cyclic && page->index > end) {
758 if (wbc->sync_mode != WB_SYNC_NONE)
759 wait_on_page_writeback(page);
761 if (PageWriteback(page) ||
762 !clear_page_dirty_for_io(page)) {
768 ret = (*writepage)(page, wbc);
769 mapping_set_error(mapping, ret);
771 bio = __mpage_writepage(bio, page, get_block,
772 &last_block_in_bio, &ret, wbc,
773 page->mapping->a_ops->writepage);
775 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
777 if (ret || (--(wbc->nr_to_write) <= 0))
779 if (wbc->nonblocking && bdi_write_congested(bdi)) {
780 wbc->encountered_congestion = 1;
784 pagevec_release(&pvec);
787 if (!scanned && !done) {
789 * We hit the last page and there is more work to be done: wrap
790 * back to the start of the file
796 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
797 mapping->writeback_index = index;
799 mpage_bio_submit(WRITE, bio);
802 EXPORT_SYMBOL(mpage_writepages);
804 int mpage_writepage(struct page *page, get_block_t get_block,
805 struct writeback_control *wbc)
809 sector_t last_block_in_bio = 0;
811 bio = __mpage_writepage(NULL, page, get_block,
812 &last_block_in_bio, &ret, wbc, NULL);
814 mpage_bio_submit(WRITE, bio);
818 EXPORT_SYMBOL(mpage_writepage);