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
);
84 end_page_writeback(page
);
85 } while (bvec
>= bio
->bi_io_vec
);
90 struct bio
*mpage_bio_submit(int rw
, struct bio
*bio
)
92 bio
->bi_end_io
= mpage_end_io_read
;
94 bio
->bi_end_io
= mpage_end_io_write
;
100 mpage_alloc(struct block_device
*bdev
,
101 sector_t first_sector
, int nr_vecs
,
102 unsigned int __nocast gfp_flags
)
106 bio
= bio_alloc(gfp_flags
, nr_vecs
);
108 if (bio
== NULL
&& (current
->flags
& PF_MEMALLOC
)) {
109 while (!bio
&& (nr_vecs
/= 2))
110 bio
= bio_alloc(gfp_flags
, nr_vecs
);
115 bio
->bi_sector
= first_sector
;
121 * support function for mpage_readpages. The fs supplied get_block might
122 * return an up to date buffer. This is used to map that buffer into
123 * the page, which allows readpage to avoid triggering a duplicate call
126 * The idea is to avoid adding buffers to pages that don't already have
127 * them. So when the buffer is up to date and the page size == block size,
128 * this marks the page up to date instead of adding new buffers.
131 map_buffer_to_page(struct page
*page
, struct buffer_head
*bh
, int page_block
)
133 struct inode
*inode
= page
->mapping
->host
;
134 struct buffer_head
*page_bh
, *head
;
137 if (!page_has_buffers(page
)) {
139 * don't make any buffers if there is only one buffer on
140 * the page and the page just needs to be set up to date
142 if (inode
->i_blkbits
== PAGE_CACHE_SHIFT
&&
143 buffer_uptodate(bh
)) {
144 SetPageUptodate(page
);
147 create_empty_buffers(page
, 1 << inode
->i_blkbits
, 0);
149 head
= page_buffers(page
);
152 if (block
== page_block
) {
153 page_bh
->b_state
= bh
->b_state
;
154 page_bh
->b_bdev
= bh
->b_bdev
;
155 page_bh
->b_blocknr
= bh
->b_blocknr
;
158 page_bh
= page_bh
->b_this_page
;
160 } while (page_bh
!= head
);
164 do_mpage_readpage(struct bio
*bio
, struct page
*page
, unsigned nr_pages
,
165 sector_t
*last_block_in_bio
, get_block_t get_block
)
167 struct inode
*inode
= page
->mapping
->host
;
168 const unsigned blkbits
= inode
->i_blkbits
;
169 const unsigned blocks_per_page
= PAGE_CACHE_SIZE
>> blkbits
;
170 const unsigned blocksize
= 1 << blkbits
;
171 sector_t block_in_file
;
173 sector_t blocks
[MAX_BUF_PER_PAGE
];
175 unsigned first_hole
= blocks_per_page
;
176 struct block_device
*bdev
= NULL
;
177 struct buffer_head bh
;
179 int fully_mapped
= 1;
181 if (page_has_buffers(page
))
184 block_in_file
= page
->index
<< (PAGE_CACHE_SHIFT
- blkbits
);
185 last_block
= (i_size_read(inode
) + blocksize
- 1) >> blkbits
;
188 for (page_block
= 0; page_block
< blocks_per_page
;
189 page_block
++, block_in_file
++) {
191 if (block_in_file
< last_block
) {
192 if (get_block(inode
, block_in_file
, &bh
, 0))
196 if (!buffer_mapped(&bh
)) {
198 if (first_hole
== blocks_per_page
)
199 first_hole
= page_block
;
203 /* some filesystems will copy data into the page during
204 * the get_block call, in which case we don't want to
205 * read it again. map_buffer_to_page copies the data
206 * we just collected from get_block into the page's buffers
207 * so readpage doesn't have to repeat the get_block call
209 if (buffer_uptodate(&bh
)) {
210 map_buffer_to_page(page
, &bh
, page_block
);
214 if (first_hole
!= blocks_per_page
)
215 goto confused
; /* hole -> non-hole */
217 /* Contiguous blocks? */
218 if (page_block
&& blocks
[page_block
-1] != bh
.b_blocknr
-1)
220 blocks
[page_block
] = bh
.b_blocknr
;
224 if (first_hole
!= blocks_per_page
) {
225 char *kaddr
= kmap_atomic(page
, KM_USER0
);
226 memset(kaddr
+ (first_hole
<< blkbits
), 0,
227 PAGE_CACHE_SIZE
- (first_hole
<< blkbits
));
228 flush_dcache_page(page
);
229 kunmap_atomic(kaddr
, KM_USER0
);
230 if (first_hole
== 0) {
231 SetPageUptodate(page
);
235 } else if (fully_mapped
) {
236 SetPageMappedToDisk(page
);
240 * This page will go to BIO. Do we need to send this BIO off first?
242 if (bio
&& (*last_block_in_bio
!= blocks
[0] - 1))
243 bio
= mpage_bio_submit(READ
, bio
);
247 bio
= mpage_alloc(bdev
, blocks
[0] << (blkbits
- 9),
248 min_t(int, nr_pages
, bio_get_nr_vecs(bdev
)),
254 length
= first_hole
<< blkbits
;
255 if (bio_add_page(bio
, page
, length
, 0) < length
) {
256 bio
= mpage_bio_submit(READ
, bio
);
260 if (buffer_boundary(&bh
) || (first_hole
!= blocks_per_page
))
261 bio
= mpage_bio_submit(READ
, bio
);
263 *last_block_in_bio
= blocks
[blocks_per_page
- 1];
269 bio
= mpage_bio_submit(READ
, bio
);
270 if (!PageUptodate(page
))
271 block_read_full_page(page
, get_block
);
278 * mpage_readpages - populate an address space with some pages, and
279 * start reads against them.
281 * @mapping: the address_space
282 * @pages: The address of a list_head which contains the target pages. These
283 * pages have their ->index populated and are otherwise uninitialised.
285 * The page at @pages->prev has the lowest file offset, and reads should be
286 * issued in @pages->prev to @pages->next order.
288 * @nr_pages: The number of pages at *@pages
289 * @get_block: The filesystem's block mapper function.
291 * This function walks the pages and the blocks within each page, building and
292 * emitting large BIOs.
294 * If anything unusual happens, such as:
296 * - encountering a page which has buffers
297 * - encountering a page which has a non-hole after a hole
298 * - encountering a page with non-contiguous blocks
300 * then this code just gives up and calls the buffer_head-based read function.
301 * It does handle a page which has holes at the end - that is a common case:
302 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
304 * BH_Boundary explanation:
306 * There is a problem. The mpage read code assembles several pages, gets all
307 * their disk mappings, and then submits them all. That's fine, but obtaining
308 * the disk mappings may require I/O. Reads of indirect blocks, for example.
310 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
311 * submitted in the following order:
312 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
313 * because the indirect block has to be read to get the mappings of blocks
314 * 13,14,15,16. Obviously, this impacts performance.
316 * So what we do it to allow the filesystem's get_block() function to set
317 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
318 * after this one will require I/O against a block which is probably close to
319 * this one. So you should push what I/O you have currently accumulated.
321 * This all causes the disk requests to be issued in the correct order.
324 mpage_readpages(struct address_space
*mapping
, struct list_head
*pages
,
325 unsigned nr_pages
, get_block_t get_block
)
327 struct bio
*bio
= NULL
;
329 sector_t last_block_in_bio
= 0;
330 struct pagevec lru_pvec
;
332 pagevec_init(&lru_pvec
, 0);
333 for (page_idx
= 0; page_idx
< nr_pages
; page_idx
++) {
334 struct page
*page
= list_entry(pages
->prev
, struct page
, lru
);
336 prefetchw(&page
->flags
);
337 list_del(&page
->lru
);
338 if (!add_to_page_cache(page
, mapping
,
339 page
->index
, GFP_KERNEL
)) {
340 bio
= do_mpage_readpage(bio
, page
,
342 &last_block_in_bio
, get_block
);
343 if (!pagevec_add(&lru_pvec
, page
))
344 __pagevec_lru_add(&lru_pvec
);
346 page_cache_release(page
);
349 pagevec_lru_add(&lru_pvec
);
350 BUG_ON(!list_empty(pages
));
352 mpage_bio_submit(READ
, bio
);
355 EXPORT_SYMBOL(mpage_readpages
);
358 * This isn't called much at all
360 int mpage_readpage(struct page
*page
, get_block_t get_block
)
362 struct bio
*bio
= NULL
;
363 sector_t last_block_in_bio
= 0;
365 bio
= do_mpage_readpage(bio
, page
, 1,
366 &last_block_in_bio
, get_block
);
368 mpage_bio_submit(READ
, bio
);
371 EXPORT_SYMBOL(mpage_readpage
);
374 * Writing is not so simple.
376 * If the page has buffers then they will be used for obtaining the disk
377 * mapping. We only support pages which are fully mapped-and-dirty, with a
378 * special case for pages which are unmapped at the end: end-of-file.
380 * If the page has no buffers (preferred) then the page is mapped here.
382 * If all blocks are found to be contiguous then the page can go into the
383 * BIO. Otherwise fall back to the mapping's writepage().
385 * FIXME: This code wants an estimate of how many pages are still to be
386 * written, so it can intelligently allocate a suitably-sized BIO. For now,
387 * just allocate full-size (16-page) BIOs.
390 __mpage_writepage(struct bio
*bio
, struct page
*page
, get_block_t get_block
,
391 sector_t
*last_block_in_bio
, int *ret
, struct writeback_control
*wbc
,
392 writepage_t writepage_fn
)
394 struct address_space
*mapping
= page
->mapping
;
395 struct inode
*inode
= page
->mapping
->host
;
396 const unsigned blkbits
= inode
->i_blkbits
;
397 unsigned long end_index
;
398 const unsigned blocks_per_page
= PAGE_CACHE_SIZE
>> blkbits
;
400 sector_t block_in_file
;
401 sector_t blocks
[MAX_BUF_PER_PAGE
];
403 unsigned first_unmapped
= blocks_per_page
;
404 struct block_device
*bdev
= NULL
;
406 sector_t boundary_block
= 0;
407 struct block_device
*boundary_bdev
= NULL
;
409 struct buffer_head map_bh
;
410 loff_t i_size
= i_size_read(inode
);
412 if (page_has_buffers(page
)) {
413 struct buffer_head
*head
= page_buffers(page
);
414 struct buffer_head
*bh
= head
;
416 /* If they're all mapped and dirty, do it */
419 BUG_ON(buffer_locked(bh
));
420 if (!buffer_mapped(bh
)) {
422 * unmapped dirty buffers are created by
423 * __set_page_dirty_buffers -> mmapped data
425 if (buffer_dirty(bh
))
427 if (first_unmapped
== blocks_per_page
)
428 first_unmapped
= page_block
;
432 if (first_unmapped
!= blocks_per_page
)
433 goto confused
; /* hole -> non-hole */
435 if (!buffer_dirty(bh
) || !buffer_uptodate(bh
))
438 if (bh
->b_blocknr
!= blocks
[page_block
-1] + 1)
441 blocks
[page_block
++] = bh
->b_blocknr
;
442 boundary
= buffer_boundary(bh
);
444 boundary_block
= bh
->b_blocknr
;
445 boundary_bdev
= bh
->b_bdev
;
448 } while ((bh
= bh
->b_this_page
) != head
);
454 * Page has buffers, but they are all unmapped. The page was
455 * created by pagein or read over a hole which was handled by
456 * block_read_full_page(). If this address_space is also
457 * using mpage_readpages then this can rarely happen.
463 * The page has no buffers: map it to disk
465 BUG_ON(!PageUptodate(page
));
466 block_in_file
= page
->index
<< (PAGE_CACHE_SHIFT
- blkbits
);
467 last_block
= (i_size
- 1) >> blkbits
;
468 map_bh
.b_page
= page
;
469 for (page_block
= 0; page_block
< blocks_per_page
; ) {
472 if (get_block(inode
, block_in_file
, &map_bh
, 1))
474 if (buffer_new(&map_bh
))
475 unmap_underlying_metadata(map_bh
.b_bdev
,
477 if (buffer_boundary(&map_bh
)) {
478 boundary_block
= map_bh
.b_blocknr
;
479 boundary_bdev
= map_bh
.b_bdev
;
482 if (map_bh
.b_blocknr
!= blocks
[page_block
-1] + 1)
485 blocks
[page_block
++] = map_bh
.b_blocknr
;
486 boundary
= buffer_boundary(&map_bh
);
487 bdev
= map_bh
.b_bdev
;
488 if (block_in_file
== last_block
)
492 BUG_ON(page_block
== 0);
494 first_unmapped
= page_block
;
497 end_index
= i_size
>> PAGE_CACHE_SHIFT
;
498 if (page
->index
>= end_index
) {
500 * The page straddles i_size. It must be zeroed out on each
501 * and every writepage invokation because it may be mmapped.
502 * "A file is mapped in multiples of the page size. For a file
503 * that is not a multiple of the page size, the remaining memory
504 * is zeroed when mapped, and writes to that region are not
505 * written out to the file."
507 unsigned offset
= i_size
& (PAGE_CACHE_SIZE
- 1);
510 if (page
->index
> end_index
|| !offset
)
512 kaddr
= kmap_atomic(page
, KM_USER0
);
513 memset(kaddr
+ offset
, 0, PAGE_CACHE_SIZE
- offset
);
514 flush_dcache_page(page
);
515 kunmap_atomic(kaddr
, KM_USER0
);
519 * This page will go to BIO. Do we need to send this BIO off first?
521 if (bio
&& *last_block_in_bio
!= blocks
[0] - 1)
522 bio
= mpage_bio_submit(WRITE
, bio
);
526 bio
= mpage_alloc(bdev
, blocks
[0] << (blkbits
- 9),
527 bio_get_nr_vecs(bdev
), GFP_NOFS
|__GFP_HIGH
);
533 * Must try to add the page before marking the buffer clean or
534 * the confused fail path above (OOM) will be very confused when
535 * it finds all bh marked clean (i.e. it will not write anything)
537 length
= first_unmapped
<< blkbits
;
538 if (bio_add_page(bio
, page
, length
, 0) < length
) {
539 bio
= mpage_bio_submit(WRITE
, bio
);
544 * OK, we have our BIO, so we can now mark the buffers clean. Make
545 * sure to only clean buffers which we know we'll be writing.
547 if (page_has_buffers(page
)) {
548 struct buffer_head
*head
= page_buffers(page
);
549 struct buffer_head
*bh
= head
;
550 unsigned buffer_counter
= 0;
553 if (buffer_counter
++ == first_unmapped
)
555 clear_buffer_dirty(bh
);
556 bh
= bh
->b_this_page
;
557 } while (bh
!= head
);
560 * we cannot drop the bh if the page is not uptodate
561 * or a concurrent readpage would fail to serialize with the bh
562 * and it would read from disk before we reach the platter.
564 if (buffer_heads_over_limit
&& PageUptodate(page
))
565 try_to_free_buffers(page
);
568 BUG_ON(PageWriteback(page
));
569 set_page_writeback(page
);
571 if (boundary
|| (first_unmapped
!= blocks_per_page
)) {
572 bio
= mpage_bio_submit(WRITE
, bio
);
573 if (boundary_block
) {
574 write_boundary_block(boundary_bdev
,
575 boundary_block
, 1 << blkbits
);
578 *last_block_in_bio
= blocks
[blocks_per_page
- 1];
584 bio
= mpage_bio_submit(WRITE
, bio
);
587 *ret
= (*writepage_fn
)(page
, wbc
);
593 * The caller has a ref on the inode, so *mapping is stable
597 set_bit(AS_ENOSPC
, &mapping
->flags
);
599 set_bit(AS_EIO
, &mapping
->flags
);
606 * mpage_writepages - walk the list of dirty pages of the given
607 * address space and writepage() all of them.
609 * @mapping: address space structure to write
610 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
611 * @get_block: the filesystem's block mapper function.
612 * If this is NULL then use a_ops->writepage. Otherwise, go
615 * This is a library function, which implements the writepages()
616 * address_space_operation.
618 * If a page is already under I/O, generic_writepages() skips it, even
619 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
620 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
621 * and msync() need to guarantee that all the data which was dirty at the time
622 * the call was made get new I/O started against them. If wbc->sync_mode is
623 * WB_SYNC_ALL then we were called for data integrity and we must wait for
624 * existing IO to complete.
627 mpage_writepages(struct address_space
*mapping
,
628 struct writeback_control
*wbc
, get_block_t get_block
)
630 return __mpage_writepages(mapping
, wbc
, get_block
,
631 mapping
->a_ops
->writepage
);
635 __mpage_writepages(struct address_space
*mapping
,
636 struct writeback_control
*wbc
, get_block_t get_block
,
637 writepage_t writepage_fn
)
639 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
640 struct bio
*bio
= NULL
;
641 sector_t last_block_in_bio
= 0;
644 int (*writepage
)(struct page
*page
, struct writeback_control
*wbc
);
648 pgoff_t end
= -1; /* Inclusive */
652 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
653 wbc
->encountered_congestion
= 1;
658 if (get_block
== NULL
)
659 writepage
= mapping
->a_ops
->writepage
;
661 pagevec_init(&pvec
, 0);
662 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
663 index
= mapping
->writeback_index
; /* Start from prev offset */
665 index
= 0; /* whole-file sweep */
668 if (wbc
->start
|| wbc
->end
) {
669 index
= wbc
->start
>> PAGE_CACHE_SHIFT
;
670 end
= wbc
->end
>> PAGE_CACHE_SHIFT
;
675 while (!done
&& (index
<= end
) &&
676 (nr_pages
= pagevec_lookup_tag(&pvec
, mapping
, &index
,
678 min(end
- index
, (pgoff_t
)PAGEVEC_SIZE
-1) + 1))) {
682 for (i
= 0; i
< nr_pages
; i
++) {
683 struct page
*page
= pvec
.pages
[i
];
686 * At this point we hold neither mapping->tree_lock nor
687 * lock on the page itself: the page may be truncated or
688 * invalidated (changing page->mapping to NULL), or even
689 * swizzled back from swapper_space to tmpfs file
695 if (unlikely(page
->mapping
!= mapping
)) {
700 if (unlikely(is_range
) && page
->index
> end
) {
706 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
707 wait_on_page_writeback(page
);
709 if (PageWriteback(page
) ||
710 !clear_page_dirty_for_io(page
)) {
716 ret
= (*writepage
)(page
, wbc
);
726 bio
= __mpage_writepage(bio
, page
, get_block
,
727 &last_block_in_bio
, &ret
, wbc
,
730 if (unlikely(ret
== WRITEPAGE_ACTIVATE
))
732 if (ret
|| (--(wbc
->nr_to_write
) <= 0))
734 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
735 wbc
->encountered_congestion
= 1;
739 pagevec_release(&pvec
);
742 if (!scanned
&& !done
) {
744 * We hit the last page and there is more work to be done: wrap
745 * back to the start of the file
752 mapping
->writeback_index
= index
;
754 mpage_bio_submit(WRITE
, bio
);
757 EXPORT_SYMBOL(mpage_writepages
);
758 EXPORT_SYMBOL(__mpage_writepages
);
760 int mpage_writepage(struct page
*page
, get_block_t get_block
,
761 struct writeback_control
*wbc
)
765 sector_t last_block_in_bio
= 0;
767 bio
= __mpage_writepage(NULL
, page
, get_block
,
768 &last_block_in_bio
, &ret
, wbc
, NULL
);
770 mpage_bio_submit(WRITE
, bio
);
774 EXPORT_SYMBOL(mpage_writepage
);