Loading...
1/*
2 * fs/mpage.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains functions related to preparing and submitting BIOs which contain
7 * multiple pagecache pages.
8 *
9 * 15May2002 Andrew Morton
10 * Initial version
11 * 27Jun2002 axboe@suse.de
12 * use bio_add_page() to build bio's just the right size
13 */
14
15#include <linux/kernel.h>
16#include <linux/module.h>
17#include <linux/mm.h>
18#include <linux/kdev_t.h>
19#include <linux/gfp.h>
20#include <linux/bio.h>
21#include <linux/fs.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>
31
32/*
33 * I/O completion handler for multipage BIOs.
34 *
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().
38 *
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.
43 */
44static void mpage_end_io(struct bio *bio, int err)
45{
46 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
47 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
48
49 do {
50 struct page *page = bvec->bv_page;
51
52 if (--bvec >= bio->bi_io_vec)
53 prefetchw(&bvec->bv_page->flags);
54 if (bio_data_dir(bio) == READ) {
55 if (uptodate) {
56 SetPageUptodate(page);
57 } else {
58 ClearPageUptodate(page);
59 SetPageError(page);
60 }
61 unlock_page(page);
62 } else { /* bio_data_dir(bio) == WRITE */
63 if (!uptodate) {
64 SetPageError(page);
65 if (page->mapping)
66 set_bit(AS_EIO, &page->mapping->flags);
67 }
68 end_page_writeback(page);
69 }
70 } while (bvec >= bio->bi_io_vec);
71 bio_put(bio);
72}
73
74static struct bio *mpage_bio_submit(int rw, struct bio *bio)
75{
76 bio->bi_end_io = mpage_end_io;
77 submit_bio(rw, bio);
78 return NULL;
79}
80
81static struct bio *
82mpage_alloc(struct block_device *bdev,
83 sector_t first_sector, int nr_vecs,
84 gfp_t gfp_flags)
85{
86 struct bio *bio;
87
88 bio = bio_alloc(gfp_flags, nr_vecs);
89
90 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
91 while (!bio && (nr_vecs /= 2))
92 bio = bio_alloc(gfp_flags, nr_vecs);
93 }
94
95 if (bio) {
96 bio->bi_bdev = bdev;
97 bio->bi_sector = first_sector;
98 }
99 return bio;
100}
101
102/*
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
106 * to get_block.
107 *
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.
111 */
112static void
113map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
114{
115 struct inode *inode = page->mapping->host;
116 struct buffer_head *page_bh, *head;
117 int block = 0;
118
119 if (!page_has_buffers(page)) {
120 /*
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
123 */
124 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
125 buffer_uptodate(bh)) {
126 SetPageUptodate(page);
127 return;
128 }
129 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
130 }
131 head = page_buffers(page);
132 page_bh = head;
133 do {
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;
138 break;
139 }
140 page_bh = page_bh->b_this_page;
141 block++;
142 } while (page_bh != head);
143}
144
145/*
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.
149 *
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
152 * get_block() call.
153 */
154static struct bio *
155do_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)
158{
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;
164 sector_t last_block;
165 sector_t last_block_in_file;
166 sector_t blocks[MAX_BUF_PER_PAGE];
167 unsigned page_block;
168 unsigned first_hole = blocks_per_page;
169 struct block_device *bdev = NULL;
170 int length;
171 int fully_mapped = 1;
172 unsigned nblocks;
173 unsigned relative_block;
174
175 if (page_has_buffers(page))
176 goto confused;
177
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;
183 page_block = 0;
184
185 /*
186 * Map blocks using the result from the previous get_blocks call first.
187 */
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;
193
194 for (relative_block = 0; ; relative_block++) {
195 if (relative_block == last) {
196 clear_buffer_mapped(map_bh);
197 break;
198 }
199 if (page_block == blocks_per_page)
200 break;
201 blocks[page_block] = map_bh->b_blocknr + map_offset +
202 relative_block;
203 page_block++;
204 block_in_file++;
205 }
206 bdev = map_bh->b_bdev;
207 }
208
209 /*
210 * Then do more get_blocks calls until we are done with this page.
211 */
212 map_bh->b_page = page;
213 while (page_block < blocks_per_page) {
214 map_bh->b_state = 0;
215 map_bh->b_size = 0;
216
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))
220 goto confused;
221 *first_logical_block = block_in_file;
222 }
223
224 if (!buffer_mapped(map_bh)) {
225 fully_mapped = 0;
226 if (first_hole == blocks_per_page)
227 first_hole = page_block;
228 page_block++;
229 block_in_file++;
230 continue;
231 }
232
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
238 */
239 if (buffer_uptodate(map_bh)) {
240 map_buffer_to_page(page, map_bh, page_block);
241 goto confused;
242 }
243
244 if (first_hole != blocks_per_page)
245 goto confused; /* hole -> non-hole */
246
247 /* Contiguous blocks? */
248 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
249 goto confused;
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);
254 break;
255 } else if (page_block == blocks_per_page)
256 break;
257 blocks[page_block] = map_bh->b_blocknr+relative_block;
258 page_block++;
259 block_in_file++;
260 }
261 bdev = map_bh->b_bdev;
262 }
263
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);
268 unlock_page(page);
269 goto out;
270 }
271 } else if (fully_mapped) {
272 SetPageMappedToDisk(page);
273 }
274
275 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
276 cleancache_get_page(page) == 0) {
277 SetPageUptodate(page);
278 goto confused;
279 }
280
281 /*
282 * This page will go to BIO. Do we need to send this BIO off first?
283 */
284 if (bio && (*last_block_in_bio != blocks[0] - 1))
285 bio = mpage_bio_submit(READ, bio);
286
287alloc_new:
288 if (bio == NULL) {
289 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
290 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
291 GFP_KERNEL);
292 if (bio == NULL)
293 goto confused;
294 }
295
296 length = first_hole << blkbits;
297 if (bio_add_page(bio, page, length, 0) < length) {
298 bio = mpage_bio_submit(READ, bio);
299 goto alloc_new;
300 }
301
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);
307 else
308 *last_block_in_bio = blocks[blocks_per_page - 1];
309out:
310 return bio;
311
312confused:
313 if (bio)
314 bio = mpage_bio_submit(READ, bio);
315 if (!PageUptodate(page))
316 block_read_full_page(page, get_block);
317 else
318 unlock_page(page);
319 goto out;
320}
321
322/**
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.
331 *
332 * This function walks the pages and the blocks within each page, building and
333 * emitting large BIOs.
334 *
335 * If anything unusual happens, such as:
336 *
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
340 *
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.
344 *
345 * BH_Boundary explanation:
346 *
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.
350 *
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
354 *
355 * because the indirect block has to be read to get the mappings of blocks
356 * 13,14,15,16. Obviously, this impacts performance.
357 *
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.
362 *
363 * This all causes the disk requests to be issued in the correct order.
364 */
365int
366mpage_readpages(struct address_space *mapping, struct list_head *pages,
367 unsigned nr_pages, get_block_t get_block)
368{
369 struct bio *bio = NULL;
370 unsigned page_idx;
371 sector_t last_block_in_bio = 0;
372 struct buffer_head map_bh;
373 unsigned long first_logical_block = 0;
374 struct blk_plug plug;
375
376 blk_start_plug(&plug);
377
378 map_bh.b_state = 0;
379 map_bh.b_size = 0;
380 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
381 struct page *page = list_entry(pages->prev, struct page, lru);
382
383 prefetchw(&page->flags);
384 list_del(&page->lru);
385 if (!add_to_page_cache_lru(page, mapping,
386 page->index, GFP_KERNEL)) {
387 bio = do_mpage_readpage(bio, page,
388 nr_pages - page_idx,
389 &last_block_in_bio, &map_bh,
390 &first_logical_block,
391 get_block);
392 }
393 page_cache_release(page);
394 }
395 BUG_ON(!list_empty(pages));
396 if (bio)
397 mpage_bio_submit(READ, bio);
398 blk_finish_plug(&plug);
399 return 0;
400}
401EXPORT_SYMBOL(mpage_readpages);
402
403/*
404 * This isn't called much at all
405 */
406int mpage_readpage(struct page *page, get_block_t get_block)
407{
408 struct bio *bio = NULL;
409 sector_t last_block_in_bio = 0;
410 struct buffer_head map_bh;
411 unsigned long first_logical_block = 0;
412
413 map_bh.b_state = 0;
414 map_bh.b_size = 0;
415 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
416 &map_bh, &first_logical_block, get_block);
417 if (bio)
418 mpage_bio_submit(READ, bio);
419 return 0;
420}
421EXPORT_SYMBOL(mpage_readpage);
422
423/*
424 * Writing is not so simple.
425 *
426 * If the page has buffers then they will be used for obtaining the disk
427 * mapping. We only support pages which are fully mapped-and-dirty, with a
428 * special case for pages which are unmapped at the end: end-of-file.
429 *
430 * If the page has no buffers (preferred) then the page is mapped here.
431 *
432 * If all blocks are found to be contiguous then the page can go into the
433 * BIO. Otherwise fall back to the mapping's writepage().
434 *
435 * FIXME: This code wants an estimate of how many pages are still to be
436 * written, so it can intelligently allocate a suitably-sized BIO. For now,
437 * just allocate full-size (16-page) BIOs.
438 */
439
440struct mpage_data {
441 struct bio *bio;
442 sector_t last_block_in_bio;
443 get_block_t *get_block;
444 unsigned use_writepage;
445};
446
447static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
448 void *data)
449{
450 struct mpage_data *mpd = data;
451 struct bio *bio = mpd->bio;
452 struct address_space *mapping = page->mapping;
453 struct inode *inode = page->mapping->host;
454 const unsigned blkbits = inode->i_blkbits;
455 unsigned long end_index;
456 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
457 sector_t last_block;
458 sector_t block_in_file;
459 sector_t blocks[MAX_BUF_PER_PAGE];
460 unsigned page_block;
461 unsigned first_unmapped = blocks_per_page;
462 struct block_device *bdev = NULL;
463 int boundary = 0;
464 sector_t boundary_block = 0;
465 struct block_device *boundary_bdev = NULL;
466 int length;
467 struct buffer_head map_bh;
468 loff_t i_size = i_size_read(inode);
469 int ret = 0;
470
471 if (page_has_buffers(page)) {
472 struct buffer_head *head = page_buffers(page);
473 struct buffer_head *bh = head;
474
475 /* If they're all mapped and dirty, do it */
476 page_block = 0;
477 do {
478 BUG_ON(buffer_locked(bh));
479 if (!buffer_mapped(bh)) {
480 /*
481 * unmapped dirty buffers are created by
482 * __set_page_dirty_buffers -> mmapped data
483 */
484 if (buffer_dirty(bh))
485 goto confused;
486 if (first_unmapped == blocks_per_page)
487 first_unmapped = page_block;
488 continue;
489 }
490
491 if (first_unmapped != blocks_per_page)
492 goto confused; /* hole -> non-hole */
493
494 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
495 goto confused;
496 if (page_block) {
497 if (bh->b_blocknr != blocks[page_block-1] + 1)
498 goto confused;
499 }
500 blocks[page_block++] = bh->b_blocknr;
501 boundary = buffer_boundary(bh);
502 if (boundary) {
503 boundary_block = bh->b_blocknr;
504 boundary_bdev = bh->b_bdev;
505 }
506 bdev = bh->b_bdev;
507 } while ((bh = bh->b_this_page) != head);
508
509 if (first_unmapped)
510 goto page_is_mapped;
511
512 /*
513 * Page has buffers, but they are all unmapped. The page was
514 * created by pagein or read over a hole which was handled by
515 * block_read_full_page(). If this address_space is also
516 * using mpage_readpages then this can rarely happen.
517 */
518 goto confused;
519 }
520
521 /*
522 * The page has no buffers: map it to disk
523 */
524 BUG_ON(!PageUptodate(page));
525 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
526 last_block = (i_size - 1) >> blkbits;
527 map_bh.b_page = page;
528 for (page_block = 0; page_block < blocks_per_page; ) {
529
530 map_bh.b_state = 0;
531 map_bh.b_size = 1 << blkbits;
532 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
533 goto confused;
534 if (buffer_new(&map_bh))
535 unmap_underlying_metadata(map_bh.b_bdev,
536 map_bh.b_blocknr);
537 if (buffer_boundary(&map_bh)) {
538 boundary_block = map_bh.b_blocknr;
539 boundary_bdev = map_bh.b_bdev;
540 }
541 if (page_block) {
542 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
543 goto confused;
544 }
545 blocks[page_block++] = map_bh.b_blocknr;
546 boundary = buffer_boundary(&map_bh);
547 bdev = map_bh.b_bdev;
548 if (block_in_file == last_block)
549 break;
550 block_in_file++;
551 }
552 BUG_ON(page_block == 0);
553
554 first_unmapped = page_block;
555
556page_is_mapped:
557 end_index = i_size >> PAGE_CACHE_SHIFT;
558 if (page->index >= end_index) {
559 /*
560 * The page straddles i_size. It must be zeroed out on each
561 * and every writepage invocation because it may be mmapped.
562 * "A file is mapped in multiples of the page size. For a file
563 * that is not a multiple of the page size, the remaining memory
564 * is zeroed when mapped, and writes to that region are not
565 * written out to the file."
566 */
567 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
568
569 if (page->index > end_index || !offset)
570 goto confused;
571 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
572 }
573
574 /*
575 * This page will go to BIO. Do we need to send this BIO off first?
576 */
577 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
578 bio = mpage_bio_submit(WRITE, bio);
579
580alloc_new:
581 if (bio == NULL) {
582 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
583 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
584 if (bio == NULL)
585 goto confused;
586 }
587
588 /*
589 * Must try to add the page before marking the buffer clean or
590 * the confused fail path above (OOM) will be very confused when
591 * it finds all bh marked clean (i.e. it will not write anything)
592 */
593 length = first_unmapped << blkbits;
594 if (bio_add_page(bio, page, length, 0) < length) {
595 bio = mpage_bio_submit(WRITE, bio);
596 goto alloc_new;
597 }
598
599 /*
600 * OK, we have our BIO, so we can now mark the buffers clean. Make
601 * sure to only clean buffers which we know we'll be writing.
602 */
603 if (page_has_buffers(page)) {
604 struct buffer_head *head = page_buffers(page);
605 struct buffer_head *bh = head;
606 unsigned buffer_counter = 0;
607
608 do {
609 if (buffer_counter++ == first_unmapped)
610 break;
611 clear_buffer_dirty(bh);
612 bh = bh->b_this_page;
613 } while (bh != head);
614
615 /*
616 * we cannot drop the bh if the page is not uptodate
617 * or a concurrent readpage would fail to serialize with the bh
618 * and it would read from disk before we reach the platter.
619 */
620 if (buffer_heads_over_limit && PageUptodate(page))
621 try_to_free_buffers(page);
622 }
623
624 BUG_ON(PageWriteback(page));
625 set_page_writeback(page);
626 unlock_page(page);
627 if (boundary || (first_unmapped != blocks_per_page)) {
628 bio = mpage_bio_submit(WRITE, bio);
629 if (boundary_block) {
630 write_boundary_block(boundary_bdev,
631 boundary_block, 1 << blkbits);
632 }
633 } else {
634 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
635 }
636 goto out;
637
638confused:
639 if (bio)
640 bio = mpage_bio_submit(WRITE, bio);
641
642 if (mpd->use_writepage) {
643 ret = mapping->a_ops->writepage(page, wbc);
644 } else {
645 ret = -EAGAIN;
646 goto out;
647 }
648 /*
649 * The caller has a ref on the inode, so *mapping is stable
650 */
651 mapping_set_error(mapping, ret);
652out:
653 mpd->bio = bio;
654 return ret;
655}
656
657/**
658 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
659 * @mapping: address space structure to write
660 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
661 * @get_block: the filesystem's block mapper function.
662 * If this is NULL then use a_ops->writepage. Otherwise, go
663 * direct-to-BIO.
664 *
665 * This is a library function, which implements the writepages()
666 * address_space_operation.
667 *
668 * If a page is already under I/O, generic_writepages() skips it, even
669 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
670 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
671 * and msync() need to guarantee that all the data which was dirty at the time
672 * the call was made get new I/O started against them. If wbc->sync_mode is
673 * WB_SYNC_ALL then we were called for data integrity and we must wait for
674 * existing IO to complete.
675 */
676int
677mpage_writepages(struct address_space *mapping,
678 struct writeback_control *wbc, get_block_t get_block)
679{
680 struct blk_plug plug;
681 int ret;
682
683 blk_start_plug(&plug);
684
685 if (!get_block)
686 ret = generic_writepages(mapping, wbc);
687 else {
688 struct mpage_data mpd = {
689 .bio = NULL,
690 .last_block_in_bio = 0,
691 .get_block = get_block,
692 .use_writepage = 1,
693 };
694
695 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
696 if (mpd.bio)
697 mpage_bio_submit(WRITE, mpd.bio);
698 }
699 blk_finish_plug(&plug);
700 return ret;
701}
702EXPORT_SYMBOL(mpage_writepages);
703
704int mpage_writepage(struct page *page, get_block_t get_block,
705 struct writeback_control *wbc)
706{
707 struct mpage_data mpd = {
708 .bio = NULL,
709 .last_block_in_bio = 0,
710 .get_block = get_block,
711 .use_writepage = 0,
712 };
713 int ret = __mpage_writepage(page, wbc, &mpd);
714 if (mpd.bio)
715 mpage_bio_submit(WRITE, mpd.bio);
716 return ret;
717}
718EXPORT_SYMBOL(mpage_writepage);
1/*
2 * fs/mpage.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains functions related to preparing and submitting BIOs which contain
7 * multiple pagecache pages.
8 *
9 * 15May2002 Andrew Morton
10 * Initial version
11 * 27Jun2002 axboe@suse.de
12 * use bio_add_page() to build bio's just the right size
13 */
14
15#include <linux/kernel.h>
16#include <linux/export.h>
17#include <linux/mm.h>
18#include <linux/kdev_t.h>
19#include <linux/gfp.h>
20#include <linux/bio.h>
21#include <linux/fs.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>
31
32/*
33 * I/O completion handler for multipage BIOs.
34 *
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().
38 *
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.
43 */
44static void mpage_end_io(struct bio *bio, int err)
45{
46 struct bio_vec *bv;
47 int i;
48
49 bio_for_each_segment_all(bv, bio, i) {
50 struct page *page = bv->bv_page;
51
52 if (bio_data_dir(bio) == READ) {
53 if (!err) {
54 SetPageUptodate(page);
55 } else {
56 ClearPageUptodate(page);
57 SetPageError(page);
58 }
59 unlock_page(page);
60 } else { /* bio_data_dir(bio) == WRITE */
61 if (err) {
62 SetPageError(page);
63 if (page->mapping)
64 set_bit(AS_EIO, &page->mapping->flags);
65 }
66 end_page_writeback(page);
67 }
68 }
69
70 bio_put(bio);
71}
72
73static struct bio *mpage_bio_submit(int rw, struct bio *bio)
74{
75 bio->bi_end_io = mpage_end_io;
76 submit_bio(rw, bio);
77 return NULL;
78}
79
80static struct bio *
81mpage_alloc(struct block_device *bdev,
82 sector_t first_sector, int nr_vecs,
83 gfp_t gfp_flags)
84{
85 struct bio *bio;
86
87 bio = bio_alloc(gfp_flags, nr_vecs);
88
89 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
90 while (!bio && (nr_vecs /= 2))
91 bio = bio_alloc(gfp_flags, nr_vecs);
92 }
93
94 if (bio) {
95 bio->bi_bdev = bdev;
96 bio->bi_iter.bi_sector = first_sector;
97 }
98 return bio;
99}
100
101/*
102 * support function for mpage_readpages. The fs supplied get_block might
103 * return an up to date buffer. This is used to map that buffer into
104 * the page, which allows readpage to avoid triggering a duplicate call
105 * to get_block.
106 *
107 * The idea is to avoid adding buffers to pages that don't already have
108 * them. So when the buffer is up to date and the page size == block size,
109 * this marks the page up to date instead of adding new buffers.
110 */
111static void
112map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
113{
114 struct inode *inode = page->mapping->host;
115 struct buffer_head *page_bh, *head;
116 int block = 0;
117
118 if (!page_has_buffers(page)) {
119 /*
120 * don't make any buffers if there is only one buffer on
121 * the page and the page just needs to be set up to date
122 */
123 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
124 buffer_uptodate(bh)) {
125 SetPageUptodate(page);
126 return;
127 }
128 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
129 }
130 head = page_buffers(page);
131 page_bh = head;
132 do {
133 if (block == page_block) {
134 page_bh->b_state = bh->b_state;
135 page_bh->b_bdev = bh->b_bdev;
136 page_bh->b_blocknr = bh->b_blocknr;
137 break;
138 }
139 page_bh = page_bh->b_this_page;
140 block++;
141 } while (page_bh != head);
142}
143
144/*
145 * This is the worker routine which does all the work of mapping the disk
146 * blocks and constructs largest possible bios, submits them for IO if the
147 * blocks are not contiguous on the disk.
148 *
149 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
150 * represent the validity of its disk mapping and to decide when to do the next
151 * get_block() call.
152 */
153static struct bio *
154do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
155 sector_t *last_block_in_bio, struct buffer_head *map_bh,
156 unsigned long *first_logical_block, get_block_t get_block)
157{
158 struct inode *inode = page->mapping->host;
159 const unsigned blkbits = inode->i_blkbits;
160 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
161 const unsigned blocksize = 1 << blkbits;
162 sector_t block_in_file;
163 sector_t last_block;
164 sector_t last_block_in_file;
165 sector_t blocks[MAX_BUF_PER_PAGE];
166 unsigned page_block;
167 unsigned first_hole = blocks_per_page;
168 struct block_device *bdev = NULL;
169 int length;
170 int fully_mapped = 1;
171 unsigned nblocks;
172 unsigned relative_block;
173
174 if (page_has_buffers(page))
175 goto confused;
176
177 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
178 last_block = block_in_file + nr_pages * blocks_per_page;
179 last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
180 if (last_block > last_block_in_file)
181 last_block = last_block_in_file;
182 page_block = 0;
183
184 /*
185 * Map blocks using the result from the previous get_blocks call first.
186 */
187 nblocks = map_bh->b_size >> blkbits;
188 if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
189 block_in_file < (*first_logical_block + nblocks)) {
190 unsigned map_offset = block_in_file - *first_logical_block;
191 unsigned last = nblocks - map_offset;
192
193 for (relative_block = 0; ; relative_block++) {
194 if (relative_block == last) {
195 clear_buffer_mapped(map_bh);
196 break;
197 }
198 if (page_block == blocks_per_page)
199 break;
200 blocks[page_block] = map_bh->b_blocknr + map_offset +
201 relative_block;
202 page_block++;
203 block_in_file++;
204 }
205 bdev = map_bh->b_bdev;
206 }
207
208 /*
209 * Then do more get_blocks calls until we are done with this page.
210 */
211 map_bh->b_page = page;
212 while (page_block < blocks_per_page) {
213 map_bh->b_state = 0;
214 map_bh->b_size = 0;
215
216 if (block_in_file < last_block) {
217 map_bh->b_size = (last_block-block_in_file) << blkbits;
218 if (get_block(inode, block_in_file, map_bh, 0))
219 goto confused;
220 *first_logical_block = block_in_file;
221 }
222
223 if (!buffer_mapped(map_bh)) {
224 fully_mapped = 0;
225 if (first_hole == blocks_per_page)
226 first_hole = page_block;
227 page_block++;
228 block_in_file++;
229 continue;
230 }
231
232 /* some filesystems will copy data into the page during
233 * the get_block call, in which case we don't want to
234 * read it again. map_buffer_to_page copies the data
235 * we just collected from get_block into the page's buffers
236 * so readpage doesn't have to repeat the get_block call
237 */
238 if (buffer_uptodate(map_bh)) {
239 map_buffer_to_page(page, map_bh, page_block);
240 goto confused;
241 }
242
243 if (first_hole != blocks_per_page)
244 goto confused; /* hole -> non-hole */
245
246 /* Contiguous blocks? */
247 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
248 goto confused;
249 nblocks = map_bh->b_size >> blkbits;
250 for (relative_block = 0; ; relative_block++) {
251 if (relative_block == nblocks) {
252 clear_buffer_mapped(map_bh);
253 break;
254 } else if (page_block == blocks_per_page)
255 break;
256 blocks[page_block] = map_bh->b_blocknr+relative_block;
257 page_block++;
258 block_in_file++;
259 }
260 bdev = map_bh->b_bdev;
261 }
262
263 if (first_hole != blocks_per_page) {
264 zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
265 if (first_hole == 0) {
266 SetPageUptodate(page);
267 unlock_page(page);
268 goto out;
269 }
270 } else if (fully_mapped) {
271 SetPageMappedToDisk(page);
272 }
273
274 if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
275 cleancache_get_page(page) == 0) {
276 SetPageUptodate(page);
277 goto confused;
278 }
279
280 /*
281 * This page will go to BIO. Do we need to send this BIO off first?
282 */
283 if (bio && (*last_block_in_bio != blocks[0] - 1))
284 bio = mpage_bio_submit(READ, bio);
285
286alloc_new:
287 if (bio == NULL) {
288 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
289 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
290 GFP_KERNEL);
291 if (bio == NULL)
292 goto confused;
293 }
294
295 length = first_hole << blkbits;
296 if (bio_add_page(bio, page, length, 0) < length) {
297 bio = mpage_bio_submit(READ, bio);
298 goto alloc_new;
299 }
300
301 relative_block = block_in_file - *first_logical_block;
302 nblocks = map_bh->b_size >> blkbits;
303 if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
304 (first_hole != blocks_per_page))
305 bio = mpage_bio_submit(READ, bio);
306 else
307 *last_block_in_bio = blocks[blocks_per_page - 1];
308out:
309 return bio;
310
311confused:
312 if (bio)
313 bio = mpage_bio_submit(READ, bio);
314 if (!PageUptodate(page))
315 block_read_full_page(page, get_block);
316 else
317 unlock_page(page);
318 goto out;
319}
320
321/**
322 * mpage_readpages - populate an address space with some pages & start reads against them
323 * @mapping: the address_space
324 * @pages: The address of a list_head which contains the target pages. These
325 * pages have their ->index populated and are otherwise uninitialised.
326 * The page at @pages->prev has the lowest file offset, and reads should be
327 * issued in @pages->prev to @pages->next order.
328 * @nr_pages: The number of pages at *@pages
329 * @get_block: The filesystem's block mapper function.
330 *
331 * This function walks the pages and the blocks within each page, building and
332 * emitting large BIOs.
333 *
334 * If anything unusual happens, such as:
335 *
336 * - encountering a page which has buffers
337 * - encountering a page which has a non-hole after a hole
338 * - encountering a page with non-contiguous blocks
339 *
340 * then this code just gives up and calls the buffer_head-based read function.
341 * It does handle a page which has holes at the end - that is a common case:
342 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
343 *
344 * BH_Boundary explanation:
345 *
346 * There is a problem. The mpage read code assembles several pages, gets all
347 * their disk mappings, and then submits them all. That's fine, but obtaining
348 * the disk mappings may require I/O. Reads of indirect blocks, for example.
349 *
350 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
351 * submitted in the following order:
352 * 12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
353 *
354 * because the indirect block has to be read to get the mappings of blocks
355 * 13,14,15,16. Obviously, this impacts performance.
356 *
357 * So what we do it to allow the filesystem's get_block() function to set
358 * BH_Boundary when it maps block 11. BH_Boundary says: mapping of the block
359 * after this one will require I/O against a block which is probably close to
360 * this one. So you should push what I/O you have currently accumulated.
361 *
362 * This all causes the disk requests to be issued in the correct order.
363 */
364int
365mpage_readpages(struct address_space *mapping, struct list_head *pages,
366 unsigned nr_pages, get_block_t get_block)
367{
368 struct bio *bio = NULL;
369 unsigned page_idx;
370 sector_t last_block_in_bio = 0;
371 struct buffer_head map_bh;
372 unsigned long first_logical_block = 0;
373
374 map_bh.b_state = 0;
375 map_bh.b_size = 0;
376 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
377 struct page *page = list_entry(pages->prev, struct page, lru);
378
379 prefetchw(&page->flags);
380 list_del(&page->lru);
381 if (!add_to_page_cache_lru(page, mapping,
382 page->index, GFP_KERNEL)) {
383 bio = do_mpage_readpage(bio, page,
384 nr_pages - page_idx,
385 &last_block_in_bio, &map_bh,
386 &first_logical_block,
387 get_block);
388 }
389 page_cache_release(page);
390 }
391 BUG_ON(!list_empty(pages));
392 if (bio)
393 mpage_bio_submit(READ, bio);
394 return 0;
395}
396EXPORT_SYMBOL(mpage_readpages);
397
398/*
399 * This isn't called much at all
400 */
401int mpage_readpage(struct page *page, get_block_t get_block)
402{
403 struct bio *bio = NULL;
404 sector_t last_block_in_bio = 0;
405 struct buffer_head map_bh;
406 unsigned long first_logical_block = 0;
407
408 map_bh.b_state = 0;
409 map_bh.b_size = 0;
410 bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
411 &map_bh, &first_logical_block, get_block);
412 if (bio)
413 mpage_bio_submit(READ, bio);
414 return 0;
415}
416EXPORT_SYMBOL(mpage_readpage);
417
418/*
419 * Writing is not so simple.
420 *
421 * If the page has buffers then they will be used for obtaining the disk
422 * mapping. We only support pages which are fully mapped-and-dirty, with a
423 * special case for pages which are unmapped at the end: end-of-file.
424 *
425 * If the page has no buffers (preferred) then the page is mapped here.
426 *
427 * If all blocks are found to be contiguous then the page can go into the
428 * BIO. Otherwise fall back to the mapping's writepage().
429 *
430 * FIXME: This code wants an estimate of how many pages are still to be
431 * written, so it can intelligently allocate a suitably-sized BIO. For now,
432 * just allocate full-size (16-page) BIOs.
433 */
434
435struct mpage_data {
436 struct bio *bio;
437 sector_t last_block_in_bio;
438 get_block_t *get_block;
439 unsigned use_writepage;
440};
441
442static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
443 void *data)
444{
445 struct mpage_data *mpd = data;
446 struct bio *bio = mpd->bio;
447 struct address_space *mapping = page->mapping;
448 struct inode *inode = page->mapping->host;
449 const unsigned blkbits = inode->i_blkbits;
450 unsigned long end_index;
451 const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
452 sector_t last_block;
453 sector_t block_in_file;
454 sector_t blocks[MAX_BUF_PER_PAGE];
455 unsigned page_block;
456 unsigned first_unmapped = blocks_per_page;
457 struct block_device *bdev = NULL;
458 int boundary = 0;
459 sector_t boundary_block = 0;
460 struct block_device *boundary_bdev = NULL;
461 int length;
462 struct buffer_head map_bh;
463 loff_t i_size = i_size_read(inode);
464 int ret = 0;
465
466 if (page_has_buffers(page)) {
467 struct buffer_head *head = page_buffers(page);
468 struct buffer_head *bh = head;
469
470 /* If they're all mapped and dirty, do it */
471 page_block = 0;
472 do {
473 BUG_ON(buffer_locked(bh));
474 if (!buffer_mapped(bh)) {
475 /*
476 * unmapped dirty buffers are created by
477 * __set_page_dirty_buffers -> mmapped data
478 */
479 if (buffer_dirty(bh))
480 goto confused;
481 if (first_unmapped == blocks_per_page)
482 first_unmapped = page_block;
483 continue;
484 }
485
486 if (first_unmapped != blocks_per_page)
487 goto confused; /* hole -> non-hole */
488
489 if (!buffer_dirty(bh) || !buffer_uptodate(bh))
490 goto confused;
491 if (page_block) {
492 if (bh->b_blocknr != blocks[page_block-1] + 1)
493 goto confused;
494 }
495 blocks[page_block++] = bh->b_blocknr;
496 boundary = buffer_boundary(bh);
497 if (boundary) {
498 boundary_block = bh->b_blocknr;
499 boundary_bdev = bh->b_bdev;
500 }
501 bdev = bh->b_bdev;
502 } while ((bh = bh->b_this_page) != head);
503
504 if (first_unmapped)
505 goto page_is_mapped;
506
507 /*
508 * Page has buffers, but they are all unmapped. The page was
509 * created by pagein or read over a hole which was handled by
510 * block_read_full_page(). If this address_space is also
511 * using mpage_readpages then this can rarely happen.
512 */
513 goto confused;
514 }
515
516 /*
517 * The page has no buffers: map it to disk
518 */
519 BUG_ON(!PageUptodate(page));
520 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
521 last_block = (i_size - 1) >> blkbits;
522 map_bh.b_page = page;
523 for (page_block = 0; page_block < blocks_per_page; ) {
524
525 map_bh.b_state = 0;
526 map_bh.b_size = 1 << blkbits;
527 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
528 goto confused;
529 if (buffer_new(&map_bh))
530 unmap_underlying_metadata(map_bh.b_bdev,
531 map_bh.b_blocknr);
532 if (buffer_boundary(&map_bh)) {
533 boundary_block = map_bh.b_blocknr;
534 boundary_bdev = map_bh.b_bdev;
535 }
536 if (page_block) {
537 if (map_bh.b_blocknr != blocks[page_block-1] + 1)
538 goto confused;
539 }
540 blocks[page_block++] = map_bh.b_blocknr;
541 boundary = buffer_boundary(&map_bh);
542 bdev = map_bh.b_bdev;
543 if (block_in_file == last_block)
544 break;
545 block_in_file++;
546 }
547 BUG_ON(page_block == 0);
548
549 first_unmapped = page_block;
550
551page_is_mapped:
552 end_index = i_size >> PAGE_CACHE_SHIFT;
553 if (page->index >= end_index) {
554 /*
555 * The page straddles i_size. It must be zeroed out on each
556 * and every writepage invocation because it may be mmapped.
557 * "A file is mapped in multiples of the page size. For a file
558 * that is not a multiple of the page size, the remaining memory
559 * is zeroed when mapped, and writes to that region are not
560 * written out to the file."
561 */
562 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
563
564 if (page->index > end_index || !offset)
565 goto confused;
566 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
567 }
568
569 /*
570 * This page will go to BIO. Do we need to send this BIO off first?
571 */
572 if (bio && mpd->last_block_in_bio != blocks[0] - 1)
573 bio = mpage_bio_submit(WRITE, bio);
574
575alloc_new:
576 if (bio == NULL) {
577 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
578 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
579 if (bio == NULL)
580 goto confused;
581 }
582
583 /*
584 * Must try to add the page before marking the buffer clean or
585 * the confused fail path above (OOM) will be very confused when
586 * it finds all bh marked clean (i.e. it will not write anything)
587 */
588 length = first_unmapped << blkbits;
589 if (bio_add_page(bio, page, length, 0) < length) {
590 bio = mpage_bio_submit(WRITE, bio);
591 goto alloc_new;
592 }
593
594 /*
595 * OK, we have our BIO, so we can now mark the buffers clean. Make
596 * sure to only clean buffers which we know we'll be writing.
597 */
598 if (page_has_buffers(page)) {
599 struct buffer_head *head = page_buffers(page);
600 struct buffer_head *bh = head;
601 unsigned buffer_counter = 0;
602
603 do {
604 if (buffer_counter++ == first_unmapped)
605 break;
606 clear_buffer_dirty(bh);
607 bh = bh->b_this_page;
608 } while (bh != head);
609
610 /*
611 * we cannot drop the bh if the page is not uptodate
612 * or a concurrent readpage would fail to serialize with the bh
613 * and it would read from disk before we reach the platter.
614 */
615 if (buffer_heads_over_limit && PageUptodate(page))
616 try_to_free_buffers(page);
617 }
618
619 BUG_ON(PageWriteback(page));
620 set_page_writeback(page);
621 unlock_page(page);
622 if (boundary || (first_unmapped != blocks_per_page)) {
623 bio = mpage_bio_submit(WRITE, bio);
624 if (boundary_block) {
625 write_boundary_block(boundary_bdev,
626 boundary_block, 1 << blkbits);
627 }
628 } else {
629 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
630 }
631 goto out;
632
633confused:
634 if (bio)
635 bio = mpage_bio_submit(WRITE, bio);
636
637 if (mpd->use_writepage) {
638 ret = mapping->a_ops->writepage(page, wbc);
639 } else {
640 ret = -EAGAIN;
641 goto out;
642 }
643 /*
644 * The caller has a ref on the inode, so *mapping is stable
645 */
646 mapping_set_error(mapping, ret);
647out:
648 mpd->bio = bio;
649 return ret;
650}
651
652/**
653 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
654 * @mapping: address space structure to write
655 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
656 * @get_block: the filesystem's block mapper function.
657 * If this is NULL then use a_ops->writepage. Otherwise, go
658 * direct-to-BIO.
659 *
660 * This is a library function, which implements the writepages()
661 * address_space_operation.
662 *
663 * If a page is already under I/O, generic_writepages() skips it, even
664 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
665 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
666 * and msync() need to guarantee that all the data which was dirty at the time
667 * the call was made get new I/O started against them. If wbc->sync_mode is
668 * WB_SYNC_ALL then we were called for data integrity and we must wait for
669 * existing IO to complete.
670 */
671int
672mpage_writepages(struct address_space *mapping,
673 struct writeback_control *wbc, get_block_t get_block)
674{
675 struct blk_plug plug;
676 int ret;
677
678 blk_start_plug(&plug);
679
680 if (!get_block)
681 ret = generic_writepages(mapping, wbc);
682 else {
683 struct mpage_data mpd = {
684 .bio = NULL,
685 .last_block_in_bio = 0,
686 .get_block = get_block,
687 .use_writepage = 1,
688 };
689
690 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
691 if (mpd.bio)
692 mpage_bio_submit(WRITE, mpd.bio);
693 }
694 blk_finish_plug(&plug);
695 return ret;
696}
697EXPORT_SYMBOL(mpage_writepages);
698
699int mpage_writepage(struct page *page, get_block_t get_block,
700 struct writeback_control *wbc)
701{
702 struct mpage_data mpd = {
703 .bio = NULL,
704 .last_block_in_bio = 0,
705 .get_block = get_block,
706 .use_writepage = 0,
707 };
708 int ret = __mpage_writepage(page, wbc, &mpd);
709 if (mpd.bio)
710 mpage_bio_submit(WRITE, mpd.bio);
711 return ret;
712}
713EXPORT_SYMBOL(mpage_writepage);