[PATCH] map multiple blocks for mpage_readpages()
[linux-2.6.git] / fs / mpage.c
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    akpm@zip.com.au
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/bio.h>
20 #include <linux/fs.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>
29
30 /*
31  * I/O completion handler for multipage BIOs.
32  *
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().
36  *
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.
41  */
42 static int mpage_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
43 {
44         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
45         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
46
47         if (bio->bi_size)
48                 return 1;
49
50         do {
51                 struct page *page = bvec->bv_page;
52
53                 if (--bvec >= bio->bi_io_vec)
54                         prefetchw(&bvec->bv_page->flags);
55
56                 if (uptodate) {
57                         SetPageUptodate(page);
58                 } else {
59                         ClearPageUptodate(page);
60                         SetPageError(page);
61                 }
62                 unlock_page(page);
63         } while (bvec >= bio->bi_io_vec);
64         bio_put(bio);
65         return 0;
66 }
67
68 static int mpage_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
69 {
70         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
71         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
72
73         if (bio->bi_size)
74                 return 1;
75
76         do {
77                 struct page *page = bvec->bv_page;
78
79                 if (--bvec >= bio->bi_io_vec)
80                         prefetchw(&bvec->bv_page->flags);
81
82                 if (!uptodate){
83                         SetPageError(page);
84                         if (page->mapping)
85                                 set_bit(AS_EIO, &page->mapping->flags);
86                 }
87                 end_page_writeback(page);
88         } while (bvec >= bio->bi_io_vec);
89         bio_put(bio);
90         return 0;
91 }
92
93 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
94 {
95         bio->bi_end_io = mpage_end_io_read;
96         if (rw == WRITE)
97                 bio->bi_end_io = mpage_end_io_write;
98         submit_bio(rw, bio);
99         return NULL;
100 }
101
102 static struct bio *
103 mpage_alloc(struct block_device *bdev,
104                 sector_t first_sector, int nr_vecs,
105                 gfp_t gfp_flags)
106 {
107         struct bio *bio;
108
109         bio = bio_alloc(gfp_flags, nr_vecs);
110
111         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
112                 while (!bio && (nr_vecs /= 2))
113                         bio = bio_alloc(gfp_flags, nr_vecs);
114         }
115
116         if (bio) {
117                 bio->bi_bdev = bdev;
118                 bio->bi_sector = first_sector;
119         }
120         return bio;
121 }
122
123 /*
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
127  * to get_block.
128  *
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.
132  */
133 static void 
134 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
135 {
136         struct inode *inode = page->mapping->host;
137         struct buffer_head *page_bh, *head;
138         int block = 0;
139
140         if (!page_has_buffers(page)) {
141                 /*
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
144                  */
145                 if (inode->i_blkbits == PAGE_CACHE_SHIFT && 
146                     buffer_uptodate(bh)) {
147                         SetPageUptodate(page);    
148                         return;
149                 }
150                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
151         }
152         head = page_buffers(page);
153         page_bh = head;
154         do {
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;
159                         break;
160                 }
161                 page_bh = page_bh->b_this_page;
162                 block++;
163         } while (page_bh != head);
164 }
165
166 /*
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.
170  *
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
173  * get_block() call.
174  */
175 static struct bio *
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)
179 {
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;
185         sector_t last_block;
186         sector_t last_block_in_file;
187         sector_t blocks[MAX_BUF_PER_PAGE];
188         unsigned page_block;
189         unsigned first_hole = blocks_per_page;
190         struct block_device *bdev = NULL;
191         int length;
192         int fully_mapped = 1;
193         unsigned nblocks;
194         unsigned relative_block;
195
196         if (page_has_buffers(page))
197                 goto confused;
198
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;
204         page_block = 0;
205
206         /*
207          * Map blocks using the result from the previous get_blocks call first.
208          */
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;
214
215                 for (relative_block = 0; ; relative_block++) {
216                         if (relative_block == last) {
217                                 clear_buffer_mapped(map_bh);
218                                 break;
219                         }
220                         if (page_block == blocks_per_page)
221                                 break;
222                         blocks[page_block] = map_bh->b_blocknr + map_offset +
223                                                 relative_block;
224                         page_block++;
225                         block_in_file++;
226                 }
227                 bdev = map_bh->b_bdev;
228         }
229
230         /*
231          * Then do more get_blocks calls until we are done with this page.
232          */
233         map_bh->b_page = page;
234         while (page_block < blocks_per_page) {
235                 map_bh->b_state = 0;
236                 map_bh->b_size = 0;
237
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))
241                                 goto confused;
242                         *first_logical_block = block_in_file;
243                 }
244
245                 if (!buffer_mapped(map_bh)) {
246                         fully_mapped = 0;
247                         if (first_hole == blocks_per_page)
248                                 first_hole = page_block;
249                         page_block++;
250                         block_in_file++;
251                         clear_buffer_mapped(map_bh);
252                         continue;
253                 }
254
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
260                  */
261                 if (buffer_uptodate(map_bh)) {
262                         map_buffer_to_page(page, map_bh, page_block);
263                         goto confused;
264                 }
265         
266                 if (first_hole != blocks_per_page)
267                         goto confused;          /* hole -> non-hole */
268
269                 /* Contiguous blocks? */
270                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
271                         goto confused;
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);
276                                 break;
277                         } else if (page_block == blocks_per_page)
278                                 break;
279                         blocks[page_block] = map_bh->b_blocknr+relative_block;
280                         page_block++;
281                         block_in_file++;
282                 }
283                 bdev = map_bh->b_bdev;
284         }
285
286         if (first_hole != blocks_per_page) {
287                 char *kaddr = kmap_atomic(page, KM_USER0);
288                 memset(kaddr + (first_hole << blkbits), 0,
289                                 PAGE_CACHE_SIZE - (first_hole << blkbits));
290                 flush_dcache_page(page);
291                 kunmap_atomic(kaddr, KM_USER0);
292                 if (first_hole == 0) {
293                         SetPageUptodate(page);
294                         unlock_page(page);
295                         goto out;
296                 }
297         } else if (fully_mapped) {
298                 SetPageMappedToDisk(page);
299         }
300
301         /*
302          * This page will go to BIO.  Do we need to send this BIO off first?
303          */
304         if (bio && (*last_block_in_bio != blocks[0] - 1))
305                 bio = mpage_bio_submit(READ, bio);
306
307 alloc_new:
308         if (bio == NULL) {
309                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
310                                 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
311                                 GFP_KERNEL);
312                 if (bio == NULL)
313                         goto confused;
314         }
315
316         length = first_hole << blkbits;
317         if (bio_add_page(bio, page, length, 0) < length) {
318                 bio = mpage_bio_submit(READ, bio);
319                 goto alloc_new;
320         }
321
322         if (buffer_boundary(map_bh) || (first_hole != blocks_per_page))
323                 bio = mpage_bio_submit(READ, bio);
324         else
325                 *last_block_in_bio = blocks[blocks_per_page - 1];
326 out:
327         return bio;
328
329 confused:
330         if (bio)
331                 bio = mpage_bio_submit(READ, bio);
332         if (!PageUptodate(page))
333                 block_read_full_page(page, get_block);
334         else
335                 unlock_page(page);
336         goto out;
337 }
338
339 /**
340  * mpage_readpages - populate an address space with some pages, and
341  *                       start reads against them.
342  *
343  * @mapping: the address_space
344  * @pages: The address of a list_head which contains the target pages.  These
345  *   pages have their ->index populated and are otherwise uninitialised.
346  *
347  *   The page at @pages->prev has the lowest file offset, and reads should be
348  *   issued in @pages->prev to @pages->next order.
349  *
350  * @nr_pages: The number of pages at *@pages
351  * @get_block: The filesystem's block mapper function.
352  *
353  * This function walks the pages and the blocks within each page, building and
354  * emitting large BIOs.
355  *
356  * If anything unusual happens, such as:
357  *
358  * - encountering a page which has buffers
359  * - encountering a page which has a non-hole after a hole
360  * - encountering a page with non-contiguous blocks
361  *
362  * then this code just gives up and calls the buffer_head-based read function.
363  * It does handle a page which has holes at the end - that is a common case:
364  * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
365  *
366  * BH_Boundary explanation:
367  *
368  * There is a problem.  The mpage read code assembles several pages, gets all
369  * their disk mappings, and then submits them all.  That's fine, but obtaining
370  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
371  *
372  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
373  * submitted in the following order:
374  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
375  * because the indirect block has to be read to get the mappings of blocks
376  * 13,14,15,16.  Obviously, this impacts performance.
377  *
378  * So what we do it to allow the filesystem's get_block() function to set
379  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
380  * after this one will require I/O against a block which is probably close to
381  * this one.  So you should push what I/O you have currently accumulated.
382  *
383  * This all causes the disk requests to be issued in the correct order.
384  */
385 int
386 mpage_readpages(struct address_space *mapping, struct list_head *pages,
387                                 unsigned nr_pages, get_block_t get_block)
388 {
389         struct bio *bio = NULL;
390         unsigned page_idx;
391         sector_t last_block_in_bio = 0;
392         struct pagevec lru_pvec;
393         struct buffer_head map_bh;
394         unsigned long first_logical_block = 0;
395
396         clear_buffer_mapped(&map_bh);
397         pagevec_init(&lru_pvec, 0);
398         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
399                 struct page *page = list_entry(pages->prev, struct page, lru);
400
401                 prefetchw(&page->flags);
402                 list_del(&page->lru);
403                 if (!add_to_page_cache(page, mapping,
404                                         page->index, GFP_KERNEL)) {
405                         bio = do_mpage_readpage(bio, page,
406                                         nr_pages - page_idx,
407                                         &last_block_in_bio, &map_bh,
408                                         &first_logical_block,
409                                         get_block);
410                         if (!pagevec_add(&lru_pvec, page))
411                                 __pagevec_lru_add(&lru_pvec);
412                 } else {
413                         page_cache_release(page);
414                 }
415         }
416         pagevec_lru_add(&lru_pvec);
417         BUG_ON(!list_empty(pages));
418         if (bio)
419                 mpage_bio_submit(READ, bio);
420         return 0;
421 }
422 EXPORT_SYMBOL(mpage_readpages);
423
424 /*
425  * This isn't called much at all
426  */
427 int mpage_readpage(struct page *page, get_block_t get_block)
428 {
429         struct bio *bio = NULL;
430         sector_t last_block_in_bio = 0;
431         struct buffer_head map_bh;
432         unsigned long first_logical_block = 0;
433
434         clear_buffer_mapped(&map_bh);
435         bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
436                         &map_bh, &first_logical_block, get_block);
437         if (bio)
438                 mpage_bio_submit(READ, bio);
439         return 0;
440 }
441 EXPORT_SYMBOL(mpage_readpage);
442
443 /*
444  * Writing is not so simple.
445  *
446  * If the page has buffers then they will be used for obtaining the disk
447  * mapping.  We only support pages which are fully mapped-and-dirty, with a
448  * special case for pages which are unmapped at the end: end-of-file.
449  *
450  * If the page has no buffers (preferred) then the page is mapped here.
451  *
452  * If all blocks are found to be contiguous then the page can go into the
453  * BIO.  Otherwise fall back to the mapping's writepage().
454  * 
455  * FIXME: This code wants an estimate of how many pages are still to be
456  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
457  * just allocate full-size (16-page) BIOs.
458  */
459 static struct bio *
460 __mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
461         sector_t *last_block_in_bio, int *ret, struct writeback_control *wbc,
462         writepage_t writepage_fn)
463 {
464         struct address_space *mapping = page->mapping;
465         struct inode *inode = page->mapping->host;
466         const unsigned blkbits = inode->i_blkbits;
467         unsigned long end_index;
468         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
469         sector_t last_block;
470         sector_t block_in_file;
471         sector_t blocks[MAX_BUF_PER_PAGE];
472         unsigned page_block;
473         unsigned first_unmapped = blocks_per_page;
474         struct block_device *bdev = NULL;
475         int boundary = 0;
476         sector_t boundary_block = 0;
477         struct block_device *boundary_bdev = NULL;
478         int length;
479         struct buffer_head map_bh;
480         loff_t i_size = i_size_read(inode);
481
482         if (page_has_buffers(page)) {
483                 struct buffer_head *head = page_buffers(page);
484                 struct buffer_head *bh = head;
485
486                 /* If they're all mapped and dirty, do it */
487                 page_block = 0;
488                 do {
489                         BUG_ON(buffer_locked(bh));
490                         if (!buffer_mapped(bh)) {
491                                 /*
492                                  * unmapped dirty buffers are created by
493                                  * __set_page_dirty_buffers -> mmapped data
494                                  */
495                                 if (buffer_dirty(bh))
496                                         goto confused;
497                                 if (first_unmapped == blocks_per_page)
498                                         first_unmapped = page_block;
499                                 continue;
500                         }
501
502                         if (first_unmapped != blocks_per_page)
503                                 goto confused;  /* hole -> non-hole */
504
505                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
506                                 goto confused;
507                         if (page_block) {
508                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
509                                         goto confused;
510                         }
511                         blocks[page_block++] = bh->b_blocknr;
512                         boundary = buffer_boundary(bh);
513                         if (boundary) {
514                                 boundary_block = bh->b_blocknr;
515                                 boundary_bdev = bh->b_bdev;
516                         }
517                         bdev = bh->b_bdev;
518                 } while ((bh = bh->b_this_page) != head);
519
520                 if (first_unmapped)
521                         goto page_is_mapped;
522
523                 /*
524                  * Page has buffers, but they are all unmapped. The page was
525                  * created by pagein or read over a hole which was handled by
526                  * block_read_full_page().  If this address_space is also
527                  * using mpage_readpages then this can rarely happen.
528                  */
529                 goto confused;
530         }
531
532         /*
533          * The page has no buffers: map it to disk
534          */
535         BUG_ON(!PageUptodate(page));
536         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
537         last_block = (i_size - 1) >> blkbits;
538         map_bh.b_page = page;
539         for (page_block = 0; page_block < blocks_per_page; ) {
540
541                 map_bh.b_state = 0;
542                 map_bh.b_size = 1 << blkbits;
543                 if (get_block(inode, block_in_file, &map_bh, 1))
544                         goto confused;
545                 if (buffer_new(&map_bh))
546                         unmap_underlying_metadata(map_bh.b_bdev,
547                                                 map_bh.b_blocknr);
548                 if (buffer_boundary(&map_bh)) {
549                         boundary_block = map_bh.b_blocknr;
550                         boundary_bdev = map_bh.b_bdev;
551                 }
552                 if (page_block) {
553                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
554                                 goto confused;
555                 }
556                 blocks[page_block++] = map_bh.b_blocknr;
557                 boundary = buffer_boundary(&map_bh);
558                 bdev = map_bh.b_bdev;
559                 if (block_in_file == last_block)
560                         break;
561                 block_in_file++;
562         }
563         BUG_ON(page_block == 0);
564
565         first_unmapped = page_block;
566
567 page_is_mapped:
568         end_index = i_size >> PAGE_CACHE_SHIFT;
569         if (page->index >= end_index) {
570                 /*
571                  * The page straddles i_size.  It must be zeroed out on each
572                  * and every writepage invokation because it may be mmapped.
573                  * "A file is mapped in multiples of the page size.  For a file
574                  * that is not a multiple of the page size, the remaining memory
575                  * is zeroed when mapped, and writes to that region are not
576                  * written out to the file."
577                  */
578                 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
579                 char *kaddr;
580
581                 if (page->index > end_index || !offset)
582                         goto confused;
583                 kaddr = kmap_atomic(page, KM_USER0);
584                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
585                 flush_dcache_page(page);
586                 kunmap_atomic(kaddr, KM_USER0);
587         }
588
589         /*
590          * This page will go to BIO.  Do we need to send this BIO off first?
591          */
592         if (bio && *last_block_in_bio != blocks[0] - 1)
593                 bio = mpage_bio_submit(WRITE, bio);
594
595 alloc_new:
596         if (bio == NULL) {
597                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
598                                 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
599                 if (bio == NULL)
600                         goto confused;
601         }
602
603         /*
604          * Must try to add the page before marking the buffer clean or
605          * the confused fail path above (OOM) will be very confused when
606          * it finds all bh marked clean (i.e. it will not write anything)
607          */
608         length = first_unmapped << blkbits;
609         if (bio_add_page(bio, page, length, 0) < length) {
610                 bio = mpage_bio_submit(WRITE, bio);
611                 goto alloc_new;
612         }
613
614         /*
615          * OK, we have our BIO, so we can now mark the buffers clean.  Make
616          * sure to only clean buffers which we know we'll be writing.
617          */
618         if (page_has_buffers(page)) {
619                 struct buffer_head *head = page_buffers(page);
620                 struct buffer_head *bh = head;
621                 unsigned buffer_counter = 0;
622
623                 do {
624                         if (buffer_counter++ == first_unmapped)
625                                 break;
626                         clear_buffer_dirty(bh);
627                         bh = bh->b_this_page;
628                 } while (bh != head);
629
630                 /*
631                  * we cannot drop the bh if the page is not uptodate
632                  * or a concurrent readpage would fail to serialize with the bh
633                  * and it would read from disk before we reach the platter.
634                  */
635                 if (buffer_heads_over_limit && PageUptodate(page))
636                         try_to_free_buffers(page);
637         }
638
639         BUG_ON(PageWriteback(page));
640         set_page_writeback(page);
641         unlock_page(page);
642         if (boundary || (first_unmapped != blocks_per_page)) {
643                 bio = mpage_bio_submit(WRITE, bio);
644                 if (boundary_block) {
645                         write_boundary_block(boundary_bdev,
646                                         boundary_block, 1 << blkbits);
647                 }
648         } else {
649                 *last_block_in_bio = blocks[blocks_per_page - 1];
650         }
651         goto out;
652
653 confused:
654         if (bio)
655                 bio = mpage_bio_submit(WRITE, bio);
656
657         if (writepage_fn) {
658                 *ret = (*writepage_fn)(page, wbc);
659         } else {
660                 *ret = -EAGAIN;
661                 goto out;
662         }
663         /*
664          * The caller has a ref on the inode, so *mapping is stable
665          */
666         if (*ret) {
667                 if (*ret == -ENOSPC)
668                         set_bit(AS_ENOSPC, &mapping->flags);
669                 else
670                         set_bit(AS_EIO, &mapping->flags);
671         }
672 out:
673         return bio;
674 }
675
676 /**
677  * mpage_writepages - walk the list of dirty pages of the given
678  * address space and writepage() all of them.
679  * 
680  * @mapping: address space structure to write
681  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
682  * @get_block: the filesystem's block mapper function.
683  *             If this is NULL then use a_ops->writepage.  Otherwise, go
684  *             direct-to-BIO.
685  *
686  * This is a library function, which implements the writepages()
687  * address_space_operation.
688  *
689  * If a page is already under I/O, generic_writepages() skips it, even
690  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
691  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
692  * and msync() need to guarantee that all the data which was dirty at the time
693  * the call was made get new I/O started against them.  If wbc->sync_mode is
694  * WB_SYNC_ALL then we were called for data integrity and we must wait for
695  * existing IO to complete.
696  */
697 int
698 mpage_writepages(struct address_space *mapping,
699                 struct writeback_control *wbc, get_block_t get_block)
700 {
701         struct backing_dev_info *bdi = mapping->backing_dev_info;
702         struct bio *bio = NULL;
703         sector_t last_block_in_bio = 0;
704         int ret = 0;
705         int done = 0;
706         int (*writepage)(struct page *page, struct writeback_control *wbc);
707         struct pagevec pvec;
708         int nr_pages;
709         pgoff_t index;
710         pgoff_t end = -1;               /* Inclusive */
711         int scanned = 0;
712         int is_range = 0;
713
714         if (wbc->nonblocking && bdi_write_congested(bdi)) {
715                 wbc->encountered_congestion = 1;
716                 return 0;
717         }
718
719         writepage = NULL;
720         if (get_block == NULL)
721                 writepage = mapping->a_ops->writepage;
722
723         pagevec_init(&pvec, 0);
724         if (wbc->sync_mode == WB_SYNC_NONE) {
725                 index = mapping->writeback_index; /* Start from prev offset */
726         } else {
727                 index = 0;                        /* whole-file sweep */
728                 scanned = 1;
729         }
730         if (wbc->start || wbc->end) {
731                 index = wbc->start >> PAGE_CACHE_SHIFT;
732                 end = wbc->end >> PAGE_CACHE_SHIFT;
733                 is_range = 1;
734                 scanned = 1;
735         }
736 retry:
737         while (!done && (index <= end) &&
738                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
739                         PAGECACHE_TAG_DIRTY,
740                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
741                 unsigned i;
742
743                 scanned = 1;
744                 for (i = 0; i < nr_pages; i++) {
745                         struct page *page = pvec.pages[i];
746
747                         /*
748                          * At this point we hold neither mapping->tree_lock nor
749                          * lock on the page itself: the page may be truncated or
750                          * invalidated (changing page->mapping to NULL), or even
751                          * swizzled back from swapper_space to tmpfs file
752                          * mapping
753                          */
754
755                         lock_page(page);
756
757                         if (unlikely(page->mapping != mapping)) {
758                                 unlock_page(page);
759                                 continue;
760                         }
761
762                         if (unlikely(is_range) && page->index > end) {
763                                 done = 1;
764                                 unlock_page(page);
765                                 continue;
766                         }
767
768                         if (wbc->sync_mode != WB_SYNC_NONE)
769                                 wait_on_page_writeback(page);
770
771                         if (PageWriteback(page) ||
772                                         !clear_page_dirty_for_io(page)) {
773                                 unlock_page(page);
774                                 continue;
775                         }
776
777                         if (writepage) {
778                                 ret = (*writepage)(page, wbc);
779                                 if (ret) {
780                                         if (ret == -ENOSPC)
781                                                 set_bit(AS_ENOSPC,
782                                                         &mapping->flags);
783                                         else
784                                                 set_bit(AS_EIO,
785                                                         &mapping->flags);
786                                 }
787                         } else {
788                                 bio = __mpage_writepage(bio, page, get_block,
789                                                 &last_block_in_bio, &ret, wbc,
790                                                 page->mapping->a_ops->writepage);
791                         }
792                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
793                                 unlock_page(page);
794                         if (ret || (--(wbc->nr_to_write) <= 0))
795                                 done = 1;
796                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
797                                 wbc->encountered_congestion = 1;
798                                 done = 1;
799                         }
800                 }
801                 pagevec_release(&pvec);
802                 cond_resched();
803         }
804         if (!scanned && !done) {
805                 /*
806                  * We hit the last page and there is more work to be done: wrap
807                  * back to the start of the file
808                  */
809                 scanned = 1;
810                 index = 0;
811                 goto retry;
812         }
813         if (!is_range)
814                 mapping->writeback_index = index;
815         if (bio)
816                 mpage_bio_submit(WRITE, bio);
817         return ret;
818 }
819 EXPORT_SYMBOL(mpage_writepages);
820
821 int mpage_writepage(struct page *page, get_block_t get_block,
822         struct writeback_control *wbc)
823 {
824         int ret = 0;
825         struct bio *bio;
826         sector_t last_block_in_bio = 0;
827
828         bio = __mpage_writepage(NULL, page, get_block,
829                         &last_block_in_bio, &ret, wbc, NULL);
830         if (bio)
831                 mpage_bio_submit(WRITE, bio);
832
833         return ret;
834 }
835 EXPORT_SYMBOL(mpage_writepage);