Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
[linux-2.6-microblaze.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    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/mm_inline.h>
28 #include <linux/writeback.h>
29 #include <linux/backing-dev.h>
30 #include <linux/pagevec.h>
31 #include <linux/cleancache.h>
32 #include "internal.h"
33
34 /*
35  * I/O completion handler for multipage BIOs.
36  *
37  * The mpage code never puts partial pages into a BIO (except for end-of-file).
38  * If a page does not map to a contiguous run of blocks then it simply falls
39  * back to block_read_full_page().
40  *
41  * Why is this?  If a page's completion depends on a number of different BIOs
42  * which can complete in any order (or at the same time) then determining the
43  * status of that page is hard.  See end_buffer_async_read() for the details.
44  * There is no point in duplicating all that complexity.
45  */
46 static void mpage_end_io(struct bio *bio)
47 {
48         struct bio_vec *bv;
49         int i;
50
51         bio_for_each_segment_all(bv, bio, i) {
52                 struct page *page = bv->bv_page;
53                 page_endio(page, op_is_write(bio_op(bio)),
54                                 blk_status_to_errno(bio->bi_status));
55         }
56
57         bio_put(bio);
58 }
59
60 static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
61 {
62         bio->bi_end_io = mpage_end_io;
63         bio_set_op_attrs(bio, op, op_flags);
64         guard_bio_eod(op, bio);
65         submit_bio(bio);
66         return NULL;
67 }
68
69 static struct bio *
70 mpage_alloc(struct block_device *bdev,
71                 sector_t first_sector, int nr_vecs,
72                 gfp_t gfp_flags)
73 {
74         struct bio *bio;
75
76         /* Restrict the given (page cache) mask for slab allocations */
77         gfp_flags &= GFP_KERNEL;
78         bio = bio_alloc(gfp_flags, nr_vecs);
79
80         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
81                 while (!bio && (nr_vecs /= 2))
82                         bio = bio_alloc(gfp_flags, nr_vecs);
83         }
84
85         if (bio) {
86                 bio_set_dev(bio, bdev);
87                 bio->bi_iter.bi_sector = first_sector;
88         }
89         return bio;
90 }
91
92 /*
93  * support function for mpage_readpages.  The fs supplied get_block might
94  * return an up to date buffer.  This is used to map that buffer into
95  * the page, which allows readpage to avoid triggering a duplicate call
96  * to get_block.
97  *
98  * The idea is to avoid adding buffers to pages that don't already have
99  * them.  So when the buffer is up to date and the page size == block size,
100  * this marks the page up to date instead of adding new buffers.
101  */
102 static void 
103 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
104 {
105         struct inode *inode = page->mapping->host;
106         struct buffer_head *page_bh, *head;
107         int block = 0;
108
109         if (!page_has_buffers(page)) {
110                 /*
111                  * don't make any buffers if there is only one buffer on
112                  * the page and the page just needs to be set up to date
113                  */
114                 if (inode->i_blkbits == PAGE_SHIFT &&
115                     buffer_uptodate(bh)) {
116                         SetPageUptodate(page);    
117                         return;
118                 }
119                 create_empty_buffers(page, i_blocksize(inode), 0);
120         }
121         head = page_buffers(page);
122         page_bh = head;
123         do {
124                 if (block == page_block) {
125                         page_bh->b_state = bh->b_state;
126                         page_bh->b_bdev = bh->b_bdev;
127                         page_bh->b_blocknr = bh->b_blocknr;
128                         break;
129                 }
130                 page_bh = page_bh->b_this_page;
131                 block++;
132         } while (page_bh != head);
133 }
134
135 /*
136  * This is the worker routine which does all the work of mapping the disk
137  * blocks and constructs largest possible bios, submits them for IO if the
138  * blocks are not contiguous on the disk.
139  *
140  * We pass a buffer_head back and forth and use its buffer_mapped() flag to
141  * represent the validity of its disk mapping and to decide when to do the next
142  * get_block() call.
143  */
144 static struct bio *
145 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
146                 sector_t *last_block_in_bio, struct buffer_head *map_bh,
147                 unsigned long *first_logical_block, get_block_t get_block,
148                 gfp_t gfp)
149 {
150         struct inode *inode = page->mapping->host;
151         const unsigned blkbits = inode->i_blkbits;
152         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
153         const unsigned blocksize = 1 << blkbits;
154         sector_t block_in_file;
155         sector_t last_block;
156         sector_t last_block_in_file;
157         sector_t blocks[MAX_BUF_PER_PAGE];
158         unsigned page_block;
159         unsigned first_hole = blocks_per_page;
160         struct block_device *bdev = NULL;
161         int length;
162         int fully_mapped = 1;
163         unsigned nblocks;
164         unsigned relative_block;
165
166         if (page_has_buffers(page))
167                 goto confused;
168
169         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
170         last_block = block_in_file + nr_pages * blocks_per_page;
171         last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
172         if (last_block > last_block_in_file)
173                 last_block = last_block_in_file;
174         page_block = 0;
175
176         /*
177          * Map blocks using the result from the previous get_blocks call first.
178          */
179         nblocks = map_bh->b_size >> blkbits;
180         if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
181                         block_in_file < (*first_logical_block + nblocks)) {
182                 unsigned map_offset = block_in_file - *first_logical_block;
183                 unsigned last = nblocks - map_offset;
184
185                 for (relative_block = 0; ; relative_block++) {
186                         if (relative_block == last) {
187                                 clear_buffer_mapped(map_bh);
188                                 break;
189                         }
190                         if (page_block == blocks_per_page)
191                                 break;
192                         blocks[page_block] = map_bh->b_blocknr + map_offset +
193                                                 relative_block;
194                         page_block++;
195                         block_in_file++;
196                 }
197                 bdev = map_bh->b_bdev;
198         }
199
200         /*
201          * Then do more get_blocks calls until we are done with this page.
202          */
203         map_bh->b_page = page;
204         while (page_block < blocks_per_page) {
205                 map_bh->b_state = 0;
206                 map_bh->b_size = 0;
207
208                 if (block_in_file < last_block) {
209                         map_bh->b_size = (last_block-block_in_file) << blkbits;
210                         if (get_block(inode, block_in_file, map_bh, 0))
211                                 goto confused;
212                         *first_logical_block = block_in_file;
213                 }
214
215                 if (!buffer_mapped(map_bh)) {
216                         fully_mapped = 0;
217                         if (first_hole == blocks_per_page)
218                                 first_hole = page_block;
219                         page_block++;
220                         block_in_file++;
221                         continue;
222                 }
223
224                 /* some filesystems will copy data into the page during
225                  * the get_block call, in which case we don't want to
226                  * read it again.  map_buffer_to_page copies the data
227                  * we just collected from get_block into the page's buffers
228                  * so readpage doesn't have to repeat the get_block call
229                  */
230                 if (buffer_uptodate(map_bh)) {
231                         map_buffer_to_page(page, map_bh, page_block);
232                         goto confused;
233                 }
234         
235                 if (first_hole != blocks_per_page)
236                         goto confused;          /* hole -> non-hole */
237
238                 /* Contiguous blocks? */
239                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
240                         goto confused;
241                 nblocks = map_bh->b_size >> blkbits;
242                 for (relative_block = 0; ; relative_block++) {
243                         if (relative_block == nblocks) {
244                                 clear_buffer_mapped(map_bh);
245                                 break;
246                         } else if (page_block == blocks_per_page)
247                                 break;
248                         blocks[page_block] = map_bh->b_blocknr+relative_block;
249                         page_block++;
250                         block_in_file++;
251                 }
252                 bdev = map_bh->b_bdev;
253         }
254
255         if (first_hole != blocks_per_page) {
256                 zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
257                 if (first_hole == 0) {
258                         SetPageUptodate(page);
259                         unlock_page(page);
260                         goto out;
261                 }
262         } else if (fully_mapped) {
263                 SetPageMappedToDisk(page);
264         }
265
266         if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
267             cleancache_get_page(page) == 0) {
268                 SetPageUptodate(page);
269                 goto confused;
270         }
271
272         /*
273          * This page will go to BIO.  Do we need to send this BIO off first?
274          */
275         if (bio && (*last_block_in_bio != blocks[0] - 1))
276                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
277
278 alloc_new:
279         if (bio == NULL) {
280                 if (first_hole == blocks_per_page) {
281                         if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
282                                                                 page))
283                                 goto out;
284                 }
285                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
286                                 min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
287                 if (bio == NULL)
288                         goto confused;
289         }
290
291         length = first_hole << blkbits;
292         if (bio_add_page(bio, page, length, 0) < length) {
293                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
294                 goto alloc_new;
295         }
296
297         relative_block = block_in_file - *first_logical_block;
298         nblocks = map_bh->b_size >> blkbits;
299         if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
300             (first_hole != blocks_per_page))
301                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
302         else
303                 *last_block_in_bio = blocks[blocks_per_page - 1];
304 out:
305         return bio;
306
307 confused:
308         if (bio)
309                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
310         if (!PageUptodate(page))
311                 block_read_full_page(page, get_block);
312         else
313                 unlock_page(page);
314         goto out;
315 }
316
317 /**
318  * mpage_readpages - populate an address space with some pages & start reads against them
319  * @mapping: the address_space
320  * @pages: The address of a list_head which contains the target pages.  These
321  *   pages have their ->index populated and are otherwise uninitialised.
322  *   The page at @pages->prev has the lowest file offset, and reads should be
323  *   issued in @pages->prev to @pages->next order.
324  * @nr_pages: The number of pages at *@pages
325  * @get_block: The filesystem's block mapper function.
326  *
327  * This function walks the pages and the blocks within each page, building and
328  * emitting large BIOs.
329  *
330  * If anything unusual happens, such as:
331  *
332  * - encountering a page which has buffers
333  * - encountering a page which has a non-hole after a hole
334  * - encountering a page with non-contiguous blocks
335  *
336  * then this code just gives up and calls the buffer_head-based read function.
337  * It does handle a page which has holes at the end - that is a common case:
338  * the end-of-file on blocksize < PAGE_SIZE setups.
339  *
340  * BH_Boundary explanation:
341  *
342  * There is a problem.  The mpage read code assembles several pages, gets all
343  * their disk mappings, and then submits them all.  That's fine, but obtaining
344  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
345  *
346  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
347  * submitted in the following order:
348  *
349  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
350  *
351  * because the indirect block has to be read to get the mappings of blocks
352  * 13,14,15,16.  Obviously, this impacts performance.
353  *
354  * So what we do it to allow the filesystem's get_block() function to set
355  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
356  * after this one will require I/O against a block which is probably close to
357  * this one.  So you should push what I/O you have currently accumulated.
358  *
359  * This all causes the disk requests to be issued in the correct order.
360  */
361 int
362 mpage_readpages(struct address_space *mapping, struct list_head *pages,
363                                 unsigned nr_pages, get_block_t get_block)
364 {
365         struct bio *bio = NULL;
366         unsigned page_idx;
367         sector_t last_block_in_bio = 0;
368         struct buffer_head map_bh;
369         unsigned long first_logical_block = 0;
370         gfp_t gfp = readahead_gfp_mask(mapping);
371
372         map_bh.b_state = 0;
373         map_bh.b_size = 0;
374         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
375                 struct page *page = lru_to_page(pages);
376
377                 prefetchw(&page->flags);
378                 list_del(&page->lru);
379                 if (!add_to_page_cache_lru(page, mapping,
380                                         page->index,
381                                         gfp)) {
382                         bio = do_mpage_readpage(bio, page,
383                                         nr_pages - page_idx,
384                                         &last_block_in_bio, &map_bh,
385                                         &first_logical_block,
386                                         get_block, gfp);
387                 }
388                 put_page(page);
389         }
390         BUG_ON(!list_empty(pages));
391         if (bio)
392                 mpage_bio_submit(REQ_OP_READ, 0, bio);
393         return 0;
394 }
395 EXPORT_SYMBOL(mpage_readpages);
396
397 /*
398  * This isn't called much at all
399  */
400 int mpage_readpage(struct page *page, get_block_t get_block)
401 {
402         struct bio *bio = NULL;
403         sector_t last_block_in_bio = 0;
404         struct buffer_head map_bh;
405         unsigned long first_logical_block = 0;
406         gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
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, gfp);
412         if (bio)
413                 mpage_bio_submit(REQ_OP_READ, 0, bio);
414         return 0;
415 }
416 EXPORT_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
435 struct 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
442 /*
443  * We have our BIO, so we can now mark the buffers clean.  Make
444  * sure to only clean buffers which we know we'll be writing.
445  */
446 static void clean_buffers(struct page *page, unsigned first_unmapped)
447 {
448         unsigned buffer_counter = 0;
449         struct buffer_head *bh, *head;
450         if (!page_has_buffers(page))
451                 return;
452         head = page_buffers(page);
453         bh = head;
454
455         do {
456                 if (buffer_counter++ == first_unmapped)
457                         break;
458                 clear_buffer_dirty(bh);
459                 bh = bh->b_this_page;
460         } while (bh != head);
461
462         /*
463          * we cannot drop the bh if the page is not uptodate or a concurrent
464          * readpage would fail to serialize with the bh and it would read from
465          * disk before we reach the platter.
466          */
467         if (buffer_heads_over_limit && PageUptodate(page))
468                 try_to_free_buffers(page);
469 }
470
471 /*
472  * For situations where we want to clean all buffers attached to a page.
473  * We don't need to calculate how many buffers are attached to the page,
474  * we just need to specify a number larger than the maximum number of buffers.
475  */
476 void clean_page_buffers(struct page *page)
477 {
478         clean_buffers(page, ~0U);
479 }
480
481 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
482                       void *data)
483 {
484         struct mpage_data *mpd = data;
485         struct bio *bio = mpd->bio;
486         struct address_space *mapping = page->mapping;
487         struct inode *inode = page->mapping->host;
488         const unsigned blkbits = inode->i_blkbits;
489         unsigned long end_index;
490         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
491         sector_t last_block;
492         sector_t block_in_file;
493         sector_t blocks[MAX_BUF_PER_PAGE];
494         unsigned page_block;
495         unsigned first_unmapped = blocks_per_page;
496         struct block_device *bdev = NULL;
497         int boundary = 0;
498         sector_t boundary_block = 0;
499         struct block_device *boundary_bdev = NULL;
500         int length;
501         struct buffer_head map_bh;
502         loff_t i_size = i_size_read(inode);
503         int ret = 0;
504         int op_flags = wbc_to_write_flags(wbc);
505
506         if (page_has_buffers(page)) {
507                 struct buffer_head *head = page_buffers(page);
508                 struct buffer_head *bh = head;
509
510                 /* If they're all mapped and dirty, do it */
511                 page_block = 0;
512                 do {
513                         BUG_ON(buffer_locked(bh));
514                         if (!buffer_mapped(bh)) {
515                                 /*
516                                  * unmapped dirty buffers are created by
517                                  * __set_page_dirty_buffers -> mmapped data
518                                  */
519                                 if (buffer_dirty(bh))
520                                         goto confused;
521                                 if (first_unmapped == blocks_per_page)
522                                         first_unmapped = page_block;
523                                 continue;
524                         }
525
526                         if (first_unmapped != blocks_per_page)
527                                 goto confused;  /* hole -> non-hole */
528
529                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
530                                 goto confused;
531                         if (page_block) {
532                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
533                                         goto confused;
534                         }
535                         blocks[page_block++] = bh->b_blocknr;
536                         boundary = buffer_boundary(bh);
537                         if (boundary) {
538                                 boundary_block = bh->b_blocknr;
539                                 boundary_bdev = bh->b_bdev;
540                         }
541                         bdev = bh->b_bdev;
542                 } while ((bh = bh->b_this_page) != head);
543
544                 if (first_unmapped)
545                         goto page_is_mapped;
546
547                 /*
548                  * Page has buffers, but they are all unmapped. The page was
549                  * created by pagein or read over a hole which was handled by
550                  * block_read_full_page().  If this address_space is also
551                  * using mpage_readpages then this can rarely happen.
552                  */
553                 goto confused;
554         }
555
556         /*
557          * The page has no buffers: map it to disk
558          */
559         BUG_ON(!PageUptodate(page));
560         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
561         last_block = (i_size - 1) >> blkbits;
562         map_bh.b_page = page;
563         for (page_block = 0; page_block < blocks_per_page; ) {
564
565                 map_bh.b_state = 0;
566                 map_bh.b_size = 1 << blkbits;
567                 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
568                         goto confused;
569                 if (buffer_new(&map_bh))
570                         clean_bdev_bh_alias(&map_bh);
571                 if (buffer_boundary(&map_bh)) {
572                         boundary_block = map_bh.b_blocknr;
573                         boundary_bdev = map_bh.b_bdev;
574                 }
575                 if (page_block) {
576                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
577                                 goto confused;
578                 }
579                 blocks[page_block++] = map_bh.b_blocknr;
580                 boundary = buffer_boundary(&map_bh);
581                 bdev = map_bh.b_bdev;
582                 if (block_in_file == last_block)
583                         break;
584                 block_in_file++;
585         }
586         BUG_ON(page_block == 0);
587
588         first_unmapped = page_block;
589
590 page_is_mapped:
591         end_index = i_size >> PAGE_SHIFT;
592         if (page->index >= end_index) {
593                 /*
594                  * The page straddles i_size.  It must be zeroed out on each
595                  * and every writepage invocation because it may be mmapped.
596                  * "A file is mapped in multiples of the page size.  For a file
597                  * that is not a multiple of the page size, the remaining memory
598                  * is zeroed when mapped, and writes to that region are not
599                  * written out to the file."
600                  */
601                 unsigned offset = i_size & (PAGE_SIZE - 1);
602
603                 if (page->index > end_index || !offset)
604                         goto confused;
605                 zero_user_segment(page, offset, PAGE_SIZE);
606         }
607
608         /*
609          * This page will go to BIO.  Do we need to send this BIO off first?
610          */
611         if (bio && mpd->last_block_in_bio != blocks[0] - 1)
612                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
613
614 alloc_new:
615         if (bio == NULL) {
616                 if (first_unmapped == blocks_per_page) {
617                         if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
618                                                                 page, wbc))
619                                 goto out;
620                 }
621                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
622                                 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
623                 if (bio == NULL)
624                         goto confused;
625
626                 wbc_init_bio(wbc, bio);
627                 bio->bi_write_hint = inode->i_write_hint;
628         }
629
630         /*
631          * Must try to add the page before marking the buffer clean or
632          * the confused fail path above (OOM) will be very confused when
633          * it finds all bh marked clean (i.e. it will not write anything)
634          */
635         wbc_account_io(wbc, page, PAGE_SIZE);
636         length = first_unmapped << blkbits;
637         if (bio_add_page(bio, page, length, 0) < length) {
638                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
639                 goto alloc_new;
640         }
641
642         clean_buffers(page, first_unmapped);
643
644         BUG_ON(PageWriteback(page));
645         set_page_writeback(page);
646         unlock_page(page);
647         if (boundary || (first_unmapped != blocks_per_page)) {
648                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
649                 if (boundary_block) {
650                         write_boundary_block(boundary_bdev,
651                                         boundary_block, 1 << blkbits);
652                 }
653         } else {
654                 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
655         }
656         goto out;
657
658 confused:
659         if (bio)
660                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
661
662         if (mpd->use_writepage) {
663                 ret = mapping->a_ops->writepage(page, wbc);
664         } else {
665                 ret = -EAGAIN;
666                 goto out;
667         }
668         /*
669          * The caller has a ref on the inode, so *mapping is stable
670          */
671         mapping_set_error(mapping, ret);
672 out:
673         mpd->bio = bio;
674         return ret;
675 }
676
677 /**
678  * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
679  * @mapping: address space structure to write
680  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
681  * @get_block: the filesystem's block mapper function.
682  *             If this is NULL then use a_ops->writepage.  Otherwise, go
683  *             direct-to-BIO.
684  *
685  * This is a library function, which implements the writepages()
686  * address_space_operation.
687  *
688  * If a page is already under I/O, generic_writepages() skips it, even
689  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
690  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
691  * and msync() need to guarantee that all the data which was dirty at the time
692  * the call was made get new I/O started against them.  If wbc->sync_mode is
693  * WB_SYNC_ALL then we were called for data integrity and we must wait for
694  * existing IO to complete.
695  */
696 int
697 mpage_writepages(struct address_space *mapping,
698                 struct writeback_control *wbc, get_block_t get_block)
699 {
700         struct blk_plug plug;
701         int ret;
702
703         blk_start_plug(&plug);
704
705         if (!get_block)
706                 ret = generic_writepages(mapping, wbc);
707         else {
708                 struct mpage_data mpd = {
709                         .bio = NULL,
710                         .last_block_in_bio = 0,
711                         .get_block = get_block,
712                         .use_writepage = 1,
713                 };
714
715                 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
716                 if (mpd.bio) {
717                         int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
718                                   REQ_SYNC : 0);
719                         mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
720                 }
721         }
722         blk_finish_plug(&plug);
723         return ret;
724 }
725 EXPORT_SYMBOL(mpage_writepages);
726
727 int mpage_writepage(struct page *page, get_block_t get_block,
728         struct writeback_control *wbc)
729 {
730         struct mpage_data mpd = {
731                 .bio = NULL,
732                 .last_block_in_bio = 0,
733                 .get_block = get_block,
734                 .use_writepage = 0,
735         };
736         int ret = __mpage_writepage(page, wbc, &mpd);
737         if (mpd.bio) {
738                 int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
739                           REQ_SYNC : 0);
740                 mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
741         }
742         return ret;
743 }
744 EXPORT_SYMBOL(mpage_writepage);