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>
17 #include <linux/kdev_t.h>
18 #include <linux/bio.h>
20 #include <linux/buffer_head.h>
21 #include <linux/blkdev.h>
22 #include <linux/highmem.h>
23 #include <linux/prefetch.h>
24 #include <linux/mpage.h>
25 #include <linux/writeback.h>
26 #include <linux/backing-dev.h>
27 #include <linux/pagevec.h>
30 * I/O completion handler for multipage BIOs.
32 * The mpage code never puts partial pages into a BIO (except for end-of-file).
33 * If a page does not map to a contiguous run of blocks then it simply falls
34 * back to block_read_full_page().
36 * Why is this? If a page's completion depends on a number of different BIOs
37 * which can complete in any order (or at the same time) then determining the
38 * status of that page is hard. See end_buffer_async_read() for the details.
39 * There is no point in duplicating all that complexity.
41 static int mpage_end_io_read(struct bio
*bio
, unsigned int bytes_done
, int err
)
43 const int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
44 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
);
56 SetPageUptodate(page
);
58 ClearPageUptodate(page
);
62 } while (bvec
>= bio
->bi_io_vec
);
67 static int mpage_end_io_write(struct bio
*bio
, unsigned int bytes_done
, int err
)
69 const int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
70 struct bio_vec
*bvec
= bio
->bi_io_vec
+ bio
->bi_vcnt
- 1;
76 struct page
*page
= bvec
->bv_page
;
78 if (--bvec
>= bio
->bi_io_vec
)
79 prefetchw(&bvec
->bv_page
->flags
);
83 end_page_writeback(page
);
84 } while (bvec
>= bio
->bi_io_vec
);
89 struct bio
*mpage_bio_submit(int rw
, struct bio
*bio
)
91 bio
->bi_end_io
= mpage_end_io_read
;
93 bio
->bi_end_io
= mpage_end_io_write
;
99 mpage_alloc(struct block_device
*bdev
,
100 sector_t first_sector
, int nr_vecs
, int gfp_flags
)
104 bio
= bio_alloc(gfp_flags
, nr_vecs
);
106 if (bio
== NULL
&& (current
->flags
& PF_MEMALLOC
)) {
107 while (!bio
&& (nr_vecs
/= 2))
108 bio
= bio_alloc(gfp_flags
, nr_vecs
);
113 bio
->bi_sector
= first_sector
;
119 * mpage_readpages - populate an address space with some pages, and
120 * start reads against them.
122 * @mapping: the address_space
123 * @pages: The address of a list_head which contains the target pages. These
124 * pages have their ->index populated and are otherwise uninitialised.
126 * The page at @pages->prev has the lowest file offset, and reads should be
127 * issued in @pages->prev to @pages->next order.
129 * @nr_pages: The number of pages at *@pages
130 * @get_block: The filesystem's block mapper function.
132 * This function walks the pages and the blocks within each page, building and
133 * emitting large BIOs.
135 * If anything unusual happens, such as:
137 * - encountering a page which has buffers
138 * - encountering a page which has a non-hole after a hole
139 * - encountering a page with non-contiguous blocks
141 * then this code just gives up and calls the buffer_head-based read function.
142 * It does handle a page which has holes at the end - that is a common case:
143 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
145 * BH_Boundary explanation:
147 * There is a problem. The mpage read code assembles several pages, gets all
148 * their disk mappings, and then submits them all. That's fine, but obtaining
149 * the disk mappings may require I/O. Reads of indirect blocks, for example.
151 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
152 * submitted in the following order:
153 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
154 * because the indirect block has to be read to get the mappings of blocks
155 * 13,14,15,16. Obviously, this impacts performance.
157 * So what we do it to allow the filesystem's get_block() function to set
158 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
159 * after this one will require I/O against a block which is probably close to
160 * this one. So you should push what I/O you have currently accumulated.
162 * This all causes the disk requests to be issued in the correct order.
165 do_mpage_readpage(struct bio
*bio
, struct page
*page
, unsigned nr_pages
,
166 sector_t
*last_block_in_bio
, get_block_t get_block
)
168 struct inode
*inode
= page
->mapping
->host
;
169 const unsigned blkbits
= inode
->i_blkbits
;
170 const unsigned blocks_per_page
= PAGE_CACHE_SIZE
>> blkbits
;
171 const unsigned blocksize
= 1 << blkbits
;
172 sector_t block_in_file
;
174 sector_t blocks
[MAX_BUF_PER_PAGE
];
176 unsigned first_hole
= blocks_per_page
;
177 struct block_device
*bdev
= NULL
;
178 struct buffer_head bh
;
181 if (page_has_buffers(page
))
184 block_in_file
= page
->index
<< (PAGE_CACHE_SHIFT
- blkbits
);
185 last_block
= (inode
->i_size
+ blocksize
- 1) >> blkbits
;
187 for (page_block
= 0; page_block
< blocks_per_page
;
188 page_block
++, block_in_file
++) {
190 if (block_in_file
< last_block
) {
191 if (get_block(inode
, block_in_file
, &bh
, 0))
195 if (!buffer_mapped(&bh
)) {
196 if (first_hole
== blocks_per_page
)
197 first_hole
= page_block
;
201 if (first_hole
!= blocks_per_page
)
202 goto confused
; /* hole -> non-hole */
204 /* Contiguous blocks? */
205 if (page_block
&& blocks
[page_block
-1] != bh
.b_blocknr
-1)
207 blocks
[page_block
] = bh
.b_blocknr
;
211 if (first_hole
!= blocks_per_page
) {
212 memset(kmap(page
) + (first_hole
<< blkbits
), 0,
213 PAGE_CACHE_SIZE
- (first_hole
<< blkbits
));
214 flush_dcache_page(page
);
216 if (first_hole
== 0) {
217 SetPageUptodate(page
);
224 * This page will go to BIO. Do we need to send this BIO off first?
226 if (bio
&& (*last_block_in_bio
!= blocks
[0] - 1))
227 bio
= mpage_bio_submit(READ
, bio
);
231 bio
= mpage_alloc(bdev
, blocks
[0] << (blkbits
- 9),
232 nr_pages
, GFP_KERNEL
);
237 length
= first_hole
<< blkbits
;
238 if (bio_add_page(bio
, page
, length
, 0) < length
) {
239 bio
= mpage_bio_submit(READ
, bio
);
243 if (buffer_boundary(&bh
) || (first_hole
!= blocks_per_page
))
244 bio
= mpage_bio_submit(READ
, bio
);
246 *last_block_in_bio
= blocks
[blocks_per_page
- 1];
252 bio
= mpage_bio_submit(READ
, bio
);
253 block_read_full_page(page
, get_block
);
258 mpage_readpages(struct address_space
*mapping
, struct list_head
*pages
,
259 unsigned nr_pages
, get_block_t get_block
)
261 struct bio
*bio
= NULL
;
263 sector_t last_block_in_bio
= 0;
264 struct pagevec lru_pvec
;
266 pagevec_init(&lru_pvec
, 0);
267 for (page_idx
= 0; page_idx
< nr_pages
; page_idx
++) {
268 struct page
*page
= list_entry(pages
->prev
, struct page
, list
);
270 prefetchw(&page
->flags
);
271 list_del(&page
->list
);
272 if (!add_to_page_cache(page
, mapping
, page
->index
)) {
273 bio
= do_mpage_readpage(bio
, page
,
275 &last_block_in_bio
, get_block
);
276 if (!pagevec_add(&lru_pvec
, page
))
277 __pagevec_lru_add(&lru_pvec
);
279 page_cache_release(page
);
282 pagevec_lru_add(&lru_pvec
);
283 BUG_ON(!list_empty(pages
));
285 mpage_bio_submit(READ
, bio
);
288 EXPORT_SYMBOL(mpage_readpages
);
291 * This isn't called much at all
293 int mpage_readpage(struct page
*page
, get_block_t get_block
)
295 struct bio
*bio
= NULL
;
296 sector_t last_block_in_bio
= 0;
298 bio
= do_mpage_readpage(bio
, page
, 1,
299 &last_block_in_bio
, get_block
);
301 mpage_bio_submit(READ
, bio
);
304 EXPORT_SYMBOL(mpage_readpage
);
307 * Writing is not so simple.
309 * If the page has buffers then they will be used for obtaining the disk
310 * mapping. We only support pages which are fully mapped-and-dirty, with a
311 * special case for pages which are unmapped at the end: end-of-file.
313 * If the page has no buffers (preferred) then the page is mapped here.
315 * If all blocks are found to be contiguous then the page can go into the
316 * BIO. Otherwise fall back to the mapping's writepage().
318 * FIXME: This code wants an estimate of how many pages are still to be
319 * written, so it can intelligently allocate a suitably-sized BIO. For now,
320 * just allocate full-size (16-page) BIOs.
323 mpage_writepage(struct bio
*bio
, struct page
*page
, get_block_t get_block
,
324 sector_t
*last_block_in_bio
, int *ret
)
326 struct inode
*inode
= page
->mapping
->host
;
327 const unsigned blkbits
= inode
->i_blkbits
;
328 unsigned long end_index
;
329 const unsigned blocks_per_page
= PAGE_CACHE_SIZE
>> blkbits
;
331 sector_t block_in_file
;
332 sector_t blocks
[MAX_BUF_PER_PAGE
];
334 unsigned first_unmapped
= blocks_per_page
;
335 struct block_device
*bdev
= NULL
;
337 sector_t boundary_block
= 0;
338 struct block_device
*boundary_bdev
= NULL
;
341 if (page_has_buffers(page
)) {
342 struct buffer_head
*head
= page_buffers(page
);
343 struct buffer_head
*bh
= head
;
345 /* If they're all mapped and dirty, do it */
348 BUG_ON(buffer_locked(bh
));
349 if (!buffer_mapped(bh
)) {
351 * unmapped dirty buffers are created by
352 * __set_page_dirty_buffers -> mmapped data
354 if (buffer_dirty(bh
))
356 if (first_unmapped
== blocks_per_page
)
357 first_unmapped
= page_block
;
361 if (first_unmapped
!= blocks_per_page
)
362 goto confused
; /* hole -> non-hole */
364 if (!buffer_dirty(bh
) || !buffer_uptodate(bh
))
367 if (bh
->b_blocknr
!= blocks
[page_block
-1] + 1)
370 blocks
[page_block
++] = bh
->b_blocknr
;
371 boundary
= buffer_boundary(bh
);
373 boundary_block
= bh
->b_blocknr
;
374 boundary_bdev
= bh
->b_bdev
;
377 } while ((bh
= bh
->b_this_page
) != head
);
383 * Page has buffers, but they are all unmapped. The page was
384 * created by pagein or read over a hole which was handled by
385 * block_read_full_page(). If this address_space is also
386 * using mpage_readpages then this can rarely happen.
392 * The page has no buffers: map it to disk
394 BUG_ON(!PageUptodate(page
));
395 block_in_file
= page
->index
<< (PAGE_CACHE_SHIFT
- blkbits
);
396 last_block
= (inode
->i_size
- 1) >> blkbits
;
397 for (page_block
= 0; page_block
< blocks_per_page
; ) {
398 struct buffer_head map_bh
;
401 if (get_block(inode
, block_in_file
, &map_bh
, 1))
403 if (buffer_new(&map_bh
))
404 unmap_underlying_metadata(map_bh
.b_bdev
,
406 if (buffer_boundary(&map_bh
)) {
407 boundary_block
= map_bh
.b_blocknr
;
408 boundary_bdev
= map_bh
.b_bdev
;
411 if (map_bh
.b_blocknr
!= blocks
[page_block
-1] + 1)
414 blocks
[page_block
++] = map_bh
.b_blocknr
;
415 boundary
= buffer_boundary(&map_bh
);
416 bdev
= map_bh
.b_bdev
;
417 if (block_in_file
== last_block
)
424 first_unmapped
= page_block
;
426 end_index
= inode
->i_size
>> PAGE_CACHE_SHIFT
;
427 if (page
->index
>= end_index
) {
428 unsigned offset
= inode
->i_size
& (PAGE_CACHE_SIZE
- 1);
430 if (page
->index
> end_index
|| !offset
)
432 memset(kmap(page
) + offset
, 0, PAGE_CACHE_SIZE
- offset
);
433 flush_dcache_page(page
);
440 * This page will go to BIO. Do we need to send this BIO off first?
442 if (bio
&& *last_block_in_bio
!= blocks
[0] - 1)
443 bio
= mpage_bio_submit(WRITE
, bio
);
447 bio
= mpage_alloc(bdev
, blocks
[0] << (blkbits
- 9),
448 bio_get_nr_vecs(bdev
), GFP_NOFS
|__GFP_HIGH
);
454 * OK, we have our BIO, so we can now mark the buffers clean. Make
455 * sure to only clean buffers which we know we'll be writing.
457 if (page_has_buffers(page
)) {
458 struct buffer_head
*head
= page_buffers(page
);
459 struct buffer_head
*bh
= head
;
460 unsigned buffer_counter
= 0;
463 if (buffer_counter
++ == first_unmapped
)
465 clear_buffer_dirty(bh
);
466 bh
= bh
->b_this_page
;
467 } while (bh
!= head
);
469 if (buffer_heads_over_limit
)
470 try_to_free_buffers(page
);
473 length
= first_unmapped
<< blkbits
;
474 if (bio_add_page(bio
, page
, length
, 0) < length
) {
475 bio
= mpage_bio_submit(WRITE
, bio
);
479 BUG_ON(PageWriteback(page
));
480 SetPageWriteback(page
);
482 if (boundary
|| (first_unmapped
!= blocks_per_page
)) {
483 bio
= mpage_bio_submit(WRITE
, bio
);
484 if (boundary_block
) {
485 write_boundary_block(boundary_bdev
,
486 boundary_block
, 1 << blkbits
);
489 *last_block_in_bio
= blocks
[blocks_per_page
- 1];
495 bio
= mpage_bio_submit(WRITE
, bio
);
496 *ret
= page
->mapping
->a_ops
->writepage(page
);
502 * mpage_writepages - walk the list of dirty pages of the given
503 * address space and writepage() all of them.
505 * @mapping: address space structure to write
506 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
507 * @get_block: the filesystem's block mapper function.
508 * If this is NULL then use a_ops->writepage. Otherwise, go
511 * This is a library function, which implements the writepages()
512 * address_space_operation.
514 * (The next two paragraphs refer to code which isn't here yet, but they
515 * explain the presence of address_space.io_pages)
517 * Pages can be moved from clean_pages or locked_pages onto dirty_pages
518 * at any time - it's not possible to lock against that. So pages which
519 * have already been added to a BIO may magically reappear on the dirty_pages
520 * list. And generic_writepages() will again try to lock those pages.
521 * But I/O has not yet been started against the page. Thus deadlock.
523 * To avoid this, the entire contents of the dirty_pages list are moved
524 * onto io_pages up-front. We then walk io_pages, locking the
525 * pages and submitting them for I/O, moving them to locked_pages.
527 * This has the added benefit of preventing a livelock which would otherwise
528 * occur if pages are being dirtied faster than we can write them out.
530 * If a page is already under I/O, generic_writepages() skips it, even
531 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
532 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
533 * and msync() need to guarentee that all the data which was dirty at the time
534 * the call was made get new I/O started against them. The way to do this is
535 * to run filemap_fdatawait() before calling filemap_fdatawrite().
537 * It's fairly rare for PageWriteback pages to be on ->dirty_pages. It
538 * means that someone redirtied the page while it was under I/O.
541 mpage_writepages(struct address_space
*mapping
,
542 struct writeback_control
*wbc
, get_block_t get_block
)
544 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
545 struct bio
*bio
= NULL
;
546 sector_t last_block_in_bio
= 0;
549 int sync
= called_for_sync();
551 int (*writepage
)(struct page
*);
553 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
555 wbc
->encountered_congestion
= 1;
560 if (get_block
== NULL
)
561 writepage
= mapping
->a_ops
->writepage
;
563 pagevec_init(&pvec
, 0);
564 write_lock(&mapping
->page_lock
);
566 list_splice_init(&mapping
->dirty_pages
, &mapping
->io_pages
);
568 while (!list_empty(&mapping
->io_pages
) && !done
) {
569 struct page
*page
= list_entry(mapping
->io_pages
.prev
,
571 list_del(&page
->list
);
572 if (PageWriteback(page
) && !sync
) {
573 if (PageDirty(page
)) {
574 list_add(&page
->list
, &mapping
->dirty_pages
);
577 list_add(&page
->list
, &mapping
->locked_pages
);
580 if (!PageDirty(page
)) {
581 list_add(&page
->list
, &mapping
->clean_pages
);
584 list_add(&page
->list
, &mapping
->locked_pages
);
586 page_cache_get(page
);
587 write_unlock(&mapping
->page_lock
);
592 wait_on_page_writeback(page
);
594 if (page
->mapping
&& !PageWriteback(page
) &&
595 test_clear_page_dirty(page
)) {
597 ret
= (*writepage
)(page
);
598 if (ret
== -EAGAIN
) {
599 __set_page_dirty_nobuffers(page
);
603 bio
= mpage_writepage(bio
, page
, get_block
,
604 &last_block_in_bio
, &ret
);
606 if ((current
->flags
& PF_MEMALLOC
) &&
607 !PageActive(page
) && PageLRU(page
)) {
608 if (!pagevec_add(&pvec
, page
))
609 pagevec_deactivate_inactive(&pvec
);
612 if (ret
|| (--(wbc
->nr_to_write
) <= 0))
614 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
616 wbc
->encountered_congestion
= 1;
624 page_cache_release(page
);
625 write_lock(&mapping
->page_lock
);
628 * Leave any remaining dirty pages on ->io_pages
630 write_unlock(&mapping
->page_lock
);
631 pagevec_deactivate_inactive(&pvec
);
633 mpage_bio_submit(WRITE
, bio
);
636 EXPORT_SYMBOL(mpage_writepages
);