Linux Audio

Check our new training course

Loading...
v4.10.11
  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 */
 46static 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)), bio->bi_error);
 54	}
 55
 56	bio_put(bio);
 57}
 58
 59static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
 60{
 61	bio->bi_end_io = mpage_end_io;
 62	bio_set_op_attrs(bio, op, op_flags);
 63	guard_bio_eod(op, bio);
 64	submit_bio(bio);
 65	return NULL;
 66}
 67
 68static struct bio *
 69mpage_alloc(struct block_device *bdev,
 70		sector_t first_sector, int nr_vecs,
 71		gfp_t gfp_flags)
 72{
 73	struct bio *bio;
 74
 75	/* Restrict the given (page cache) mask for slab allocations */
 76	gfp_flags &= GFP_KERNEL;
 77	bio = bio_alloc(gfp_flags, nr_vecs);
 78
 79	if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 80		while (!bio && (nr_vecs /= 2))
 81			bio = bio_alloc(gfp_flags, nr_vecs);
 82	}
 83
 84	if (bio) {
 85		bio->bi_bdev = bdev;
 86		bio->bi_iter.bi_sector = first_sector;
 87	}
 88	return bio;
 89}
 90
 91/*
 92 * support function for mpage_readpages.  The fs supplied get_block might
 93 * return an up to date buffer.  This is used to map that buffer into
 94 * the page, which allows readpage to avoid triggering a duplicate call
 95 * to get_block.
 96 *
 97 * The idea is to avoid adding buffers to pages that don't already have
 98 * them.  So when the buffer is up to date and the page size == block size,
 99 * this marks the page up to date instead of adding new buffers.
100 */
101static void 
102map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
103{
104	struct inode *inode = page->mapping->host;
105	struct buffer_head *page_bh, *head;
106	int block = 0;
107
108	if (!page_has_buffers(page)) {
109		/*
110		 * don't make any buffers if there is only one buffer on
111		 * the page and the page just needs to be set up to date
112		 */
113		if (inode->i_blkbits == PAGE_SHIFT &&
114		    buffer_uptodate(bh)) {
115			SetPageUptodate(page);    
116			return;
117		}
118		create_empty_buffers(page, 1 << inode->i_blkbits, 0);
119	}
120	head = page_buffers(page);
121	page_bh = head;
122	do {
123		if (block == page_block) {
124			page_bh->b_state = bh->b_state;
125			page_bh->b_bdev = bh->b_bdev;
126			page_bh->b_blocknr = bh->b_blocknr;
127			break;
128		}
129		page_bh = page_bh->b_this_page;
130		block++;
131	} while (page_bh != head);
132}
133
134/*
135 * This is the worker routine which does all the work of mapping the disk
136 * blocks and constructs largest possible bios, submits them for IO if the
137 * blocks are not contiguous on the disk.
138 *
139 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
140 * represent the validity of its disk mapping and to decide when to do the next
141 * get_block() call.
142 */
143static struct bio *
144do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
145		sector_t *last_block_in_bio, struct buffer_head *map_bh,
146		unsigned long *first_logical_block, get_block_t get_block,
147		gfp_t gfp)
148{
149	struct inode *inode = page->mapping->host;
150	const unsigned blkbits = inode->i_blkbits;
151	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
152	const unsigned blocksize = 1 << blkbits;
153	sector_t block_in_file;
154	sector_t last_block;
155	sector_t last_block_in_file;
156	sector_t blocks[MAX_BUF_PER_PAGE];
157	unsigned page_block;
158	unsigned first_hole = blocks_per_page;
159	struct block_device *bdev = NULL;
160	int length;
161	int fully_mapped = 1;
162	unsigned nblocks;
163	unsigned relative_block;
164
165	if (page_has_buffers(page))
166		goto confused;
167
168	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
169	last_block = block_in_file + nr_pages * blocks_per_page;
170	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
171	if (last_block > last_block_in_file)
172		last_block = last_block_in_file;
173	page_block = 0;
174
175	/*
176	 * Map blocks using the result from the previous get_blocks call first.
177	 */
178	nblocks = map_bh->b_size >> blkbits;
179	if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
180			block_in_file < (*first_logical_block + nblocks)) {
181		unsigned map_offset = block_in_file - *first_logical_block;
182		unsigned last = nblocks - map_offset;
183
184		for (relative_block = 0; ; relative_block++) {
185			if (relative_block == last) {
186				clear_buffer_mapped(map_bh);
187				break;
188			}
189			if (page_block == blocks_per_page)
190				break;
191			blocks[page_block] = map_bh->b_blocknr + map_offset +
192						relative_block;
193			page_block++;
194			block_in_file++;
195		}
196		bdev = map_bh->b_bdev;
197	}
198
199	/*
200	 * Then do more get_blocks calls until we are done with this page.
201	 */
202	map_bh->b_page = page;
203	while (page_block < blocks_per_page) {
204		map_bh->b_state = 0;
205		map_bh->b_size = 0;
206
207		if (block_in_file < last_block) {
208			map_bh->b_size = (last_block-block_in_file) << blkbits;
209			if (get_block(inode, block_in_file, map_bh, 0))
210				goto confused;
211			*first_logical_block = block_in_file;
212		}
213
214		if (!buffer_mapped(map_bh)) {
215			fully_mapped = 0;
216			if (first_hole == blocks_per_page)
217				first_hole = page_block;
218			page_block++;
219			block_in_file++;
220			continue;
221		}
222
223		/* some filesystems will copy data into the page during
224		 * the get_block call, in which case we don't want to
225		 * read it again.  map_buffer_to_page copies the data
226		 * we just collected from get_block into the page's buffers
227		 * so readpage doesn't have to repeat the get_block call
228		 */
229		if (buffer_uptodate(map_bh)) {
230			map_buffer_to_page(page, map_bh, page_block);
231			goto confused;
232		}
233	
234		if (first_hole != blocks_per_page)
235			goto confused;		/* hole -> non-hole */
236
237		/* Contiguous blocks? */
238		if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
239			goto confused;
240		nblocks = map_bh->b_size >> blkbits;
241		for (relative_block = 0; ; relative_block++) {
242			if (relative_block == nblocks) {
243				clear_buffer_mapped(map_bh);
244				break;
245			} else if (page_block == blocks_per_page)
246				break;
247			blocks[page_block] = map_bh->b_blocknr+relative_block;
248			page_block++;
249			block_in_file++;
250		}
251		bdev = map_bh->b_bdev;
252	}
253
254	if (first_hole != blocks_per_page) {
255		zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
256		if (first_hole == 0) {
257			SetPageUptodate(page);
258			unlock_page(page);
259			goto out;
260		}
261	} else if (fully_mapped) {
262		SetPageMappedToDisk(page);
263	}
264
265	if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
266	    cleancache_get_page(page) == 0) {
267		SetPageUptodate(page);
268		goto confused;
269	}
270
271	/*
272	 * This page will go to BIO.  Do we need to send this BIO off first?
273	 */
274	if (bio && (*last_block_in_bio != blocks[0] - 1))
275		bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
276
277alloc_new:
278	if (bio == NULL) {
279		if (first_hole == blocks_per_page) {
280			if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
281								page))
282				goto out;
283		}
284		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
285				min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
286		if (bio == NULL)
287			goto confused;
288	}
289
290	length = first_hole << blkbits;
291	if (bio_add_page(bio, page, length, 0) < length) {
292		bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
293		goto alloc_new;
294	}
295
296	relative_block = block_in_file - *first_logical_block;
297	nblocks = map_bh->b_size >> blkbits;
298	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
299	    (first_hole != blocks_per_page))
300		bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
301	else
302		*last_block_in_bio = blocks[blocks_per_page - 1];
303out:
304	return bio;
305
306confused:
307	if (bio)
308		bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
309	if (!PageUptodate(page))
310	        block_read_full_page(page, get_block);
311	else
312		unlock_page(page);
313	goto out;
314}
315
316/**
317 * mpage_readpages - populate an address space with some pages & start reads against them
318 * @mapping: the address_space
319 * @pages: The address of a list_head which contains the target pages.  These
320 *   pages have their ->index populated and are otherwise uninitialised.
321 *   The page at @pages->prev has the lowest file offset, and reads should be
322 *   issued in @pages->prev to @pages->next order.
323 * @nr_pages: The number of pages at *@pages
324 * @get_block: The filesystem's block mapper function.
325 *
326 * This function walks the pages and the blocks within each page, building and
327 * emitting large BIOs.
328 *
329 * If anything unusual happens, such as:
330 *
331 * - encountering a page which has buffers
332 * - encountering a page which has a non-hole after a hole
333 * - encountering a page with non-contiguous blocks
334 *
335 * then this code just gives up and calls the buffer_head-based read function.
336 * It does handle a page which has holes at the end - that is a common case:
337 * the end-of-file on blocksize < PAGE_SIZE setups.
338 *
339 * BH_Boundary explanation:
340 *
341 * There is a problem.  The mpage read code assembles several pages, gets all
342 * their disk mappings, and then submits them all.  That's fine, but obtaining
343 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
344 *
345 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
346 * submitted in the following order:
347 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
348 *
349 * because the indirect block has to be read to get the mappings of blocks
350 * 13,14,15,16.  Obviously, this impacts performance.
351 *
352 * So what we do it to allow the filesystem's get_block() function to set
353 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
354 * after this one will require I/O against a block which is probably close to
355 * this one.  So you should push what I/O you have currently accumulated.
356 *
357 * This all causes the disk requests to be issued in the correct order.
358 */
359int
360mpage_readpages(struct address_space *mapping, struct list_head *pages,
361				unsigned nr_pages, get_block_t get_block)
362{
363	struct bio *bio = NULL;
364	unsigned page_idx;
365	sector_t last_block_in_bio = 0;
366	struct buffer_head map_bh;
367	unsigned long first_logical_block = 0;
368	gfp_t gfp = readahead_gfp_mask(mapping);
369
370	map_bh.b_state = 0;
371	map_bh.b_size = 0;
372	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
373		struct page *page = lru_to_page(pages);
374
375		prefetchw(&page->flags);
376		list_del(&page->lru);
377		if (!add_to_page_cache_lru(page, mapping,
378					page->index,
379					gfp)) {
380			bio = do_mpage_readpage(bio, page,
381					nr_pages - page_idx,
382					&last_block_in_bio, &map_bh,
383					&first_logical_block,
384					get_block, gfp);
385		}
386		put_page(page);
387	}
388	BUG_ON(!list_empty(pages));
389	if (bio)
390		mpage_bio_submit(REQ_OP_READ, 0, bio);
391	return 0;
392}
393EXPORT_SYMBOL(mpage_readpages);
394
395/*
396 * This isn't called much at all
397 */
398int mpage_readpage(struct page *page, get_block_t get_block)
399{
400	struct bio *bio = NULL;
401	sector_t last_block_in_bio = 0;
402	struct buffer_head map_bh;
403	unsigned long first_logical_block = 0;
404	gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
405
406	map_bh.b_state = 0;
407	map_bh.b_size = 0;
408	bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
409			&map_bh, &first_logical_block, get_block, gfp);
410	if (bio)
411		mpage_bio_submit(REQ_OP_READ, 0, bio);
412	return 0;
413}
414EXPORT_SYMBOL(mpage_readpage);
415
416/*
417 * Writing is not so simple.
418 *
419 * If the page has buffers then they will be used for obtaining the disk
420 * mapping.  We only support pages which are fully mapped-and-dirty, with a
421 * special case for pages which are unmapped at the end: end-of-file.
422 *
423 * If the page has no buffers (preferred) then the page is mapped here.
424 *
425 * If all blocks are found to be contiguous then the page can go into the
426 * BIO.  Otherwise fall back to the mapping's writepage().
427 * 
428 * FIXME: This code wants an estimate of how many pages are still to be
429 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
430 * just allocate full-size (16-page) BIOs.
431 */
432
433struct mpage_data {
434	struct bio *bio;
435	sector_t last_block_in_bio;
436	get_block_t *get_block;
437	unsigned use_writepage;
438};
439
440/*
441 * We have our BIO, so we can now mark the buffers clean.  Make
442 * sure to only clean buffers which we know we'll be writing.
443 */
444static void clean_buffers(struct page *page, unsigned first_unmapped)
445{
446	unsigned buffer_counter = 0;
447	struct buffer_head *bh, *head;
448	if (!page_has_buffers(page))
449		return;
450	head = page_buffers(page);
451	bh = head;
452
453	do {
454		if (buffer_counter++ == first_unmapped)
455			break;
456		clear_buffer_dirty(bh);
457		bh = bh->b_this_page;
458	} while (bh != head);
459
460	/*
461	 * we cannot drop the bh if the page is not uptodate or a concurrent
462	 * readpage would fail to serialize with the bh and it would read from
463	 * disk before we reach the platter.
464	 */
465	if (buffer_heads_over_limit && PageUptodate(page))
466		try_to_free_buffers(page);
467}
468
469static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
470		      void *data)
471{
472	struct mpage_data *mpd = data;
473	struct bio *bio = mpd->bio;
474	struct address_space *mapping = page->mapping;
475	struct inode *inode = page->mapping->host;
476	const unsigned blkbits = inode->i_blkbits;
477	unsigned long end_index;
478	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
479	sector_t last_block;
480	sector_t block_in_file;
481	sector_t blocks[MAX_BUF_PER_PAGE];
482	unsigned page_block;
483	unsigned first_unmapped = blocks_per_page;
484	struct block_device *bdev = NULL;
485	int boundary = 0;
486	sector_t boundary_block = 0;
487	struct block_device *boundary_bdev = NULL;
488	int length;
489	struct buffer_head map_bh;
490	loff_t i_size = i_size_read(inode);
491	int ret = 0;
492	int op_flags = wbc_to_write_flags(wbc);
493
494	if (page_has_buffers(page)) {
495		struct buffer_head *head = page_buffers(page);
496		struct buffer_head *bh = head;
497
498		/* If they're all mapped and dirty, do it */
499		page_block = 0;
500		do {
501			BUG_ON(buffer_locked(bh));
502			if (!buffer_mapped(bh)) {
503				/*
504				 * unmapped dirty buffers are created by
505				 * __set_page_dirty_buffers -> mmapped data
506				 */
507				if (buffer_dirty(bh))
508					goto confused;
509				if (first_unmapped == blocks_per_page)
510					first_unmapped = page_block;
511				continue;
512			}
513
514			if (first_unmapped != blocks_per_page)
515				goto confused;	/* hole -> non-hole */
516
517			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
518				goto confused;
519			if (page_block) {
520				if (bh->b_blocknr != blocks[page_block-1] + 1)
521					goto confused;
522			}
523			blocks[page_block++] = bh->b_blocknr;
524			boundary = buffer_boundary(bh);
525			if (boundary) {
526				boundary_block = bh->b_blocknr;
527				boundary_bdev = bh->b_bdev;
528			}
529			bdev = bh->b_bdev;
530		} while ((bh = bh->b_this_page) != head);
531
532		if (first_unmapped)
533			goto page_is_mapped;
534
535		/*
536		 * Page has buffers, but they are all unmapped. The page was
537		 * created by pagein or read over a hole which was handled by
538		 * block_read_full_page().  If this address_space is also
539		 * using mpage_readpages then this can rarely happen.
540		 */
541		goto confused;
542	}
543
544	/*
545	 * The page has no buffers: map it to disk
546	 */
547	BUG_ON(!PageUptodate(page));
548	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
549	last_block = (i_size - 1) >> blkbits;
550	map_bh.b_page = page;
551	for (page_block = 0; page_block < blocks_per_page; ) {
552
553		map_bh.b_state = 0;
554		map_bh.b_size = 1 << blkbits;
555		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
556			goto confused;
557		if (buffer_new(&map_bh))
558			clean_bdev_bh_alias(&map_bh);
 
559		if (buffer_boundary(&map_bh)) {
560			boundary_block = map_bh.b_blocknr;
561			boundary_bdev = map_bh.b_bdev;
562		}
563		if (page_block) {
564			if (map_bh.b_blocknr != blocks[page_block-1] + 1)
565				goto confused;
566		}
567		blocks[page_block++] = map_bh.b_blocknr;
568		boundary = buffer_boundary(&map_bh);
569		bdev = map_bh.b_bdev;
570		if (block_in_file == last_block)
571			break;
572		block_in_file++;
573	}
574	BUG_ON(page_block == 0);
575
576	first_unmapped = page_block;
577
578page_is_mapped:
579	end_index = i_size >> PAGE_SHIFT;
580	if (page->index >= end_index) {
581		/*
582		 * The page straddles i_size.  It must be zeroed out on each
583		 * and every writepage invocation because it may be mmapped.
584		 * "A file is mapped in multiples of the page size.  For a file
585		 * that is not a multiple of the page size, the remaining memory
586		 * is zeroed when mapped, and writes to that region are not
587		 * written out to the file."
588		 */
589		unsigned offset = i_size & (PAGE_SIZE - 1);
590
591		if (page->index > end_index || !offset)
592			goto confused;
593		zero_user_segment(page, offset, PAGE_SIZE);
594	}
595
596	/*
597	 * This page will go to BIO.  Do we need to send this BIO off first?
598	 */
599	if (bio && mpd->last_block_in_bio != blocks[0] - 1)
600		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
601
602alloc_new:
603	if (bio == NULL) {
604		if (first_unmapped == blocks_per_page) {
605			if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
606								page, wbc)) {
607				clean_buffers(page, first_unmapped);
608				goto out;
609			}
610		}
611		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
612				BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
613		if (bio == NULL)
614			goto confused;
615
616		wbc_init_bio(wbc, bio);
617	}
618
619	/*
620	 * Must try to add the page before marking the buffer clean or
621	 * the confused fail path above (OOM) will be very confused when
622	 * it finds all bh marked clean (i.e. it will not write anything)
623	 */
624	wbc_account_io(wbc, page, PAGE_SIZE);
625	length = first_unmapped << blkbits;
626	if (bio_add_page(bio, page, length, 0) < length) {
627		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
628		goto alloc_new;
629	}
630
631	clean_buffers(page, first_unmapped);
632
633	BUG_ON(PageWriteback(page));
634	set_page_writeback(page);
635	unlock_page(page);
636	if (boundary || (first_unmapped != blocks_per_page)) {
637		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
638		if (boundary_block) {
639			write_boundary_block(boundary_bdev,
640					boundary_block, 1 << blkbits);
641		}
642	} else {
643		mpd->last_block_in_bio = blocks[blocks_per_page - 1];
644	}
645	goto out;
646
647confused:
648	if (bio)
649		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
650
651	if (mpd->use_writepage) {
652		ret = mapping->a_ops->writepage(page, wbc);
653	} else {
654		ret = -EAGAIN;
655		goto out;
656	}
657	/*
658	 * The caller has a ref on the inode, so *mapping is stable
659	 */
660	mapping_set_error(mapping, ret);
661out:
662	mpd->bio = bio;
663	return ret;
664}
665
666/**
667 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
668 * @mapping: address space structure to write
669 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
670 * @get_block: the filesystem's block mapper function.
671 *             If this is NULL then use a_ops->writepage.  Otherwise, go
672 *             direct-to-BIO.
673 *
674 * This is a library function, which implements the writepages()
675 * address_space_operation.
676 *
677 * If a page is already under I/O, generic_writepages() skips it, even
678 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
679 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
680 * and msync() need to guarantee that all the data which was dirty at the time
681 * the call was made get new I/O started against them.  If wbc->sync_mode is
682 * WB_SYNC_ALL then we were called for data integrity and we must wait for
683 * existing IO to complete.
684 */
685int
686mpage_writepages(struct address_space *mapping,
687		struct writeback_control *wbc, get_block_t get_block)
688{
689	struct blk_plug plug;
690	int ret;
691
692	blk_start_plug(&plug);
693
694	if (!get_block)
695		ret = generic_writepages(mapping, wbc);
696	else {
697		struct mpage_data mpd = {
698			.bio = NULL,
699			.last_block_in_bio = 0,
700			.get_block = get_block,
701			.use_writepage = 1,
702		};
703
704		ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
705		if (mpd.bio) {
706			int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
707				  REQ_SYNC : 0);
708			mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
709		}
710	}
711	blk_finish_plug(&plug);
712	return ret;
713}
714EXPORT_SYMBOL(mpage_writepages);
715
716int mpage_writepage(struct page *page, get_block_t get_block,
717	struct writeback_control *wbc)
718{
719	struct mpage_data mpd = {
720		.bio = NULL,
721		.last_block_in_bio = 0,
722		.get_block = get_block,
723		.use_writepage = 0,
724	};
725	int ret = __mpage_writepage(page, wbc, &mpd);
726	if (mpd.bio) {
727		int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
728			  REQ_SYNC : 0);
729		mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
730	}
731	return ret;
732}
733EXPORT_SYMBOL(mpage_writepage);
v4.6
  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 */
 46static 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, bio_data_dir(bio), bio->bi_error);
 54	}
 55
 56	bio_put(bio);
 57}
 58
 59static struct bio *mpage_bio_submit(int rw, struct bio *bio)
 60{
 61	bio->bi_end_io = mpage_end_io;
 62	guard_bio_eod(rw, bio);
 63	submit_bio(rw, bio);
 
 64	return NULL;
 65}
 66
 67static struct bio *
 68mpage_alloc(struct block_device *bdev,
 69		sector_t first_sector, int nr_vecs,
 70		gfp_t gfp_flags)
 71{
 72	struct bio *bio;
 73
 
 
 74	bio = bio_alloc(gfp_flags, nr_vecs);
 75
 76	if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 77		while (!bio && (nr_vecs /= 2))
 78			bio = bio_alloc(gfp_flags, nr_vecs);
 79	}
 80
 81	if (bio) {
 82		bio->bi_bdev = bdev;
 83		bio->bi_iter.bi_sector = first_sector;
 84	}
 85	return bio;
 86}
 87
 88/*
 89 * support function for mpage_readpages.  The fs supplied get_block might
 90 * return an up to date buffer.  This is used to map that buffer into
 91 * the page, which allows readpage to avoid triggering a duplicate call
 92 * to get_block.
 93 *
 94 * The idea is to avoid adding buffers to pages that don't already have
 95 * them.  So when the buffer is up to date and the page size == block size,
 96 * this marks the page up to date instead of adding new buffers.
 97 */
 98static void 
 99map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
100{
101	struct inode *inode = page->mapping->host;
102	struct buffer_head *page_bh, *head;
103	int block = 0;
104
105	if (!page_has_buffers(page)) {
106		/*
107		 * don't make any buffers if there is only one buffer on
108		 * the page and the page just needs to be set up to date
109		 */
110		if (inode->i_blkbits == PAGE_SHIFT &&
111		    buffer_uptodate(bh)) {
112			SetPageUptodate(page);    
113			return;
114		}
115		create_empty_buffers(page, 1 << inode->i_blkbits, 0);
116	}
117	head = page_buffers(page);
118	page_bh = head;
119	do {
120		if (block == page_block) {
121			page_bh->b_state = bh->b_state;
122			page_bh->b_bdev = bh->b_bdev;
123			page_bh->b_blocknr = bh->b_blocknr;
124			break;
125		}
126		page_bh = page_bh->b_this_page;
127		block++;
128	} while (page_bh != head);
129}
130
131/*
132 * This is the worker routine which does all the work of mapping the disk
133 * blocks and constructs largest possible bios, submits them for IO if the
134 * blocks are not contiguous on the disk.
135 *
136 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
137 * represent the validity of its disk mapping and to decide when to do the next
138 * get_block() call.
139 */
140static struct bio *
141do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
142		sector_t *last_block_in_bio, struct buffer_head *map_bh,
143		unsigned long *first_logical_block, get_block_t get_block,
144		gfp_t gfp)
145{
146	struct inode *inode = page->mapping->host;
147	const unsigned blkbits = inode->i_blkbits;
148	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
149	const unsigned blocksize = 1 << blkbits;
150	sector_t block_in_file;
151	sector_t last_block;
152	sector_t last_block_in_file;
153	sector_t blocks[MAX_BUF_PER_PAGE];
154	unsigned page_block;
155	unsigned first_hole = blocks_per_page;
156	struct block_device *bdev = NULL;
157	int length;
158	int fully_mapped = 1;
159	unsigned nblocks;
160	unsigned relative_block;
161
162	if (page_has_buffers(page))
163		goto confused;
164
165	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
166	last_block = block_in_file + nr_pages * blocks_per_page;
167	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
168	if (last_block > last_block_in_file)
169		last_block = last_block_in_file;
170	page_block = 0;
171
172	/*
173	 * Map blocks using the result from the previous get_blocks call first.
174	 */
175	nblocks = map_bh->b_size >> blkbits;
176	if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
177			block_in_file < (*first_logical_block + nblocks)) {
178		unsigned map_offset = block_in_file - *first_logical_block;
179		unsigned last = nblocks - map_offset;
180
181		for (relative_block = 0; ; relative_block++) {
182			if (relative_block == last) {
183				clear_buffer_mapped(map_bh);
184				break;
185			}
186			if (page_block == blocks_per_page)
187				break;
188			blocks[page_block] = map_bh->b_blocknr + map_offset +
189						relative_block;
190			page_block++;
191			block_in_file++;
192		}
193		bdev = map_bh->b_bdev;
194	}
195
196	/*
197	 * Then do more get_blocks calls until we are done with this page.
198	 */
199	map_bh->b_page = page;
200	while (page_block < blocks_per_page) {
201		map_bh->b_state = 0;
202		map_bh->b_size = 0;
203
204		if (block_in_file < last_block) {
205			map_bh->b_size = (last_block-block_in_file) << blkbits;
206			if (get_block(inode, block_in_file, map_bh, 0))
207				goto confused;
208			*first_logical_block = block_in_file;
209		}
210
211		if (!buffer_mapped(map_bh)) {
212			fully_mapped = 0;
213			if (first_hole == blocks_per_page)
214				first_hole = page_block;
215			page_block++;
216			block_in_file++;
217			continue;
218		}
219
220		/* some filesystems will copy data into the page during
221		 * the get_block call, in which case we don't want to
222		 * read it again.  map_buffer_to_page copies the data
223		 * we just collected from get_block into the page's buffers
224		 * so readpage doesn't have to repeat the get_block call
225		 */
226		if (buffer_uptodate(map_bh)) {
227			map_buffer_to_page(page, map_bh, page_block);
228			goto confused;
229		}
230	
231		if (first_hole != blocks_per_page)
232			goto confused;		/* hole -> non-hole */
233
234		/* Contiguous blocks? */
235		if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
236			goto confused;
237		nblocks = map_bh->b_size >> blkbits;
238		for (relative_block = 0; ; relative_block++) {
239			if (relative_block == nblocks) {
240				clear_buffer_mapped(map_bh);
241				break;
242			} else if (page_block == blocks_per_page)
243				break;
244			blocks[page_block] = map_bh->b_blocknr+relative_block;
245			page_block++;
246			block_in_file++;
247		}
248		bdev = map_bh->b_bdev;
249	}
250
251	if (first_hole != blocks_per_page) {
252		zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
253		if (first_hole == 0) {
254			SetPageUptodate(page);
255			unlock_page(page);
256			goto out;
257		}
258	} else if (fully_mapped) {
259		SetPageMappedToDisk(page);
260	}
261
262	if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
263	    cleancache_get_page(page) == 0) {
264		SetPageUptodate(page);
265		goto confused;
266	}
267
268	/*
269	 * This page will go to BIO.  Do we need to send this BIO off first?
270	 */
271	if (bio && (*last_block_in_bio != blocks[0] - 1))
272		bio = mpage_bio_submit(READ, bio);
273
274alloc_new:
275	if (bio == NULL) {
276		if (first_hole == blocks_per_page) {
277			if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
278								page))
279				goto out;
280		}
281		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
282				min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
283		if (bio == NULL)
284			goto confused;
285	}
286
287	length = first_hole << blkbits;
288	if (bio_add_page(bio, page, length, 0) < length) {
289		bio = mpage_bio_submit(READ, bio);
290		goto alloc_new;
291	}
292
293	relative_block = block_in_file - *first_logical_block;
294	nblocks = map_bh->b_size >> blkbits;
295	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
296	    (first_hole != blocks_per_page))
297		bio = mpage_bio_submit(READ, bio);
298	else
299		*last_block_in_bio = blocks[blocks_per_page - 1];
300out:
301	return bio;
302
303confused:
304	if (bio)
305		bio = mpage_bio_submit(READ, bio);
306	if (!PageUptodate(page))
307	        block_read_full_page(page, get_block);
308	else
309		unlock_page(page);
310	goto out;
311}
312
313/**
314 * mpage_readpages - populate an address space with some pages & start reads against them
315 * @mapping: the address_space
316 * @pages: The address of a list_head which contains the target pages.  These
317 *   pages have their ->index populated and are otherwise uninitialised.
318 *   The page at @pages->prev has the lowest file offset, and reads should be
319 *   issued in @pages->prev to @pages->next order.
320 * @nr_pages: The number of pages at *@pages
321 * @get_block: The filesystem's block mapper function.
322 *
323 * This function walks the pages and the blocks within each page, building and
324 * emitting large BIOs.
325 *
326 * If anything unusual happens, such as:
327 *
328 * - encountering a page which has buffers
329 * - encountering a page which has a non-hole after a hole
330 * - encountering a page with non-contiguous blocks
331 *
332 * then this code just gives up and calls the buffer_head-based read function.
333 * It does handle a page which has holes at the end - that is a common case:
334 * the end-of-file on blocksize < PAGE_SIZE setups.
335 *
336 * BH_Boundary explanation:
337 *
338 * There is a problem.  The mpage read code assembles several pages, gets all
339 * their disk mappings, and then submits them all.  That's fine, but obtaining
340 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
341 *
342 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
343 * submitted in the following order:
344 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
345 *
346 * because the indirect block has to be read to get the mappings of blocks
347 * 13,14,15,16.  Obviously, this impacts performance.
348 *
349 * So what we do it to allow the filesystem's get_block() function to set
350 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
351 * after this one will require I/O against a block which is probably close to
352 * this one.  So you should push what I/O you have currently accumulated.
353 *
354 * This all causes the disk requests to be issued in the correct order.
355 */
356int
357mpage_readpages(struct address_space *mapping, struct list_head *pages,
358				unsigned nr_pages, get_block_t get_block)
359{
360	struct bio *bio = NULL;
361	unsigned page_idx;
362	sector_t last_block_in_bio = 0;
363	struct buffer_head map_bh;
364	unsigned long first_logical_block = 0;
365	gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
366
367	map_bh.b_state = 0;
368	map_bh.b_size = 0;
369	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
370		struct page *page = lru_to_page(pages);
371
372		prefetchw(&page->flags);
373		list_del(&page->lru);
374		if (!add_to_page_cache_lru(page, mapping,
375					page->index,
376					gfp)) {
377			bio = do_mpage_readpage(bio, page,
378					nr_pages - page_idx,
379					&last_block_in_bio, &map_bh,
380					&first_logical_block,
381					get_block, gfp);
382		}
383		put_page(page);
384	}
385	BUG_ON(!list_empty(pages));
386	if (bio)
387		mpage_bio_submit(READ, bio);
388	return 0;
389}
390EXPORT_SYMBOL(mpage_readpages);
391
392/*
393 * This isn't called much at all
394 */
395int mpage_readpage(struct page *page, get_block_t get_block)
396{
397	struct bio *bio = NULL;
398	sector_t last_block_in_bio = 0;
399	struct buffer_head map_bh;
400	unsigned long first_logical_block = 0;
401	gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
402
403	map_bh.b_state = 0;
404	map_bh.b_size = 0;
405	bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
406			&map_bh, &first_logical_block, get_block, gfp);
407	if (bio)
408		mpage_bio_submit(READ, bio);
409	return 0;
410}
411EXPORT_SYMBOL(mpage_readpage);
412
413/*
414 * Writing is not so simple.
415 *
416 * If the page has buffers then they will be used for obtaining the disk
417 * mapping.  We only support pages which are fully mapped-and-dirty, with a
418 * special case for pages which are unmapped at the end: end-of-file.
419 *
420 * If the page has no buffers (preferred) then the page is mapped here.
421 *
422 * If all blocks are found to be contiguous then the page can go into the
423 * BIO.  Otherwise fall back to the mapping's writepage().
424 * 
425 * FIXME: This code wants an estimate of how many pages are still to be
426 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
427 * just allocate full-size (16-page) BIOs.
428 */
429
430struct mpage_data {
431	struct bio *bio;
432	sector_t last_block_in_bio;
433	get_block_t *get_block;
434	unsigned use_writepage;
435};
436
437/*
438 * We have our BIO, so we can now mark the buffers clean.  Make
439 * sure to only clean buffers which we know we'll be writing.
440 */
441static void clean_buffers(struct page *page, unsigned first_unmapped)
442{
443	unsigned buffer_counter = 0;
444	struct buffer_head *bh, *head;
445	if (!page_has_buffers(page))
446		return;
447	head = page_buffers(page);
448	bh = head;
449
450	do {
451		if (buffer_counter++ == first_unmapped)
452			break;
453		clear_buffer_dirty(bh);
454		bh = bh->b_this_page;
455	} while (bh != head);
456
457	/*
458	 * we cannot drop the bh if the page is not uptodate or a concurrent
459	 * readpage would fail to serialize with the bh and it would read from
460	 * disk before we reach the platter.
461	 */
462	if (buffer_heads_over_limit && PageUptodate(page))
463		try_to_free_buffers(page);
464}
465
466static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
467		      void *data)
468{
469	struct mpage_data *mpd = data;
470	struct bio *bio = mpd->bio;
471	struct address_space *mapping = page->mapping;
472	struct inode *inode = page->mapping->host;
473	const unsigned blkbits = inode->i_blkbits;
474	unsigned long end_index;
475	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
476	sector_t last_block;
477	sector_t block_in_file;
478	sector_t blocks[MAX_BUF_PER_PAGE];
479	unsigned page_block;
480	unsigned first_unmapped = blocks_per_page;
481	struct block_device *bdev = NULL;
482	int boundary = 0;
483	sector_t boundary_block = 0;
484	struct block_device *boundary_bdev = NULL;
485	int length;
486	struct buffer_head map_bh;
487	loff_t i_size = i_size_read(inode);
488	int ret = 0;
489	int wr = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
490
491	if (page_has_buffers(page)) {
492		struct buffer_head *head = page_buffers(page);
493		struct buffer_head *bh = head;
494
495		/* If they're all mapped and dirty, do it */
496		page_block = 0;
497		do {
498			BUG_ON(buffer_locked(bh));
499			if (!buffer_mapped(bh)) {
500				/*
501				 * unmapped dirty buffers are created by
502				 * __set_page_dirty_buffers -> mmapped data
503				 */
504				if (buffer_dirty(bh))
505					goto confused;
506				if (first_unmapped == blocks_per_page)
507					first_unmapped = page_block;
508				continue;
509			}
510
511			if (first_unmapped != blocks_per_page)
512				goto confused;	/* hole -> non-hole */
513
514			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
515				goto confused;
516			if (page_block) {
517				if (bh->b_blocknr != blocks[page_block-1] + 1)
518					goto confused;
519			}
520			blocks[page_block++] = bh->b_blocknr;
521			boundary = buffer_boundary(bh);
522			if (boundary) {
523				boundary_block = bh->b_blocknr;
524				boundary_bdev = bh->b_bdev;
525			}
526			bdev = bh->b_bdev;
527		} while ((bh = bh->b_this_page) != head);
528
529		if (first_unmapped)
530			goto page_is_mapped;
531
532		/*
533		 * Page has buffers, but they are all unmapped. The page was
534		 * created by pagein or read over a hole which was handled by
535		 * block_read_full_page().  If this address_space is also
536		 * using mpage_readpages then this can rarely happen.
537		 */
538		goto confused;
539	}
540
541	/*
542	 * The page has no buffers: map it to disk
543	 */
544	BUG_ON(!PageUptodate(page));
545	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
546	last_block = (i_size - 1) >> blkbits;
547	map_bh.b_page = page;
548	for (page_block = 0; page_block < blocks_per_page; ) {
549
550		map_bh.b_state = 0;
551		map_bh.b_size = 1 << blkbits;
552		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
553			goto confused;
554		if (buffer_new(&map_bh))
555			unmap_underlying_metadata(map_bh.b_bdev,
556						map_bh.b_blocknr);
557		if (buffer_boundary(&map_bh)) {
558			boundary_block = map_bh.b_blocknr;
559			boundary_bdev = map_bh.b_bdev;
560		}
561		if (page_block) {
562			if (map_bh.b_blocknr != blocks[page_block-1] + 1)
563				goto confused;
564		}
565		blocks[page_block++] = map_bh.b_blocknr;
566		boundary = buffer_boundary(&map_bh);
567		bdev = map_bh.b_bdev;
568		if (block_in_file == last_block)
569			break;
570		block_in_file++;
571	}
572	BUG_ON(page_block == 0);
573
574	first_unmapped = page_block;
575
576page_is_mapped:
577	end_index = i_size >> PAGE_SHIFT;
578	if (page->index >= end_index) {
579		/*
580		 * The page straddles i_size.  It must be zeroed out on each
581		 * and every writepage invocation because it may be mmapped.
582		 * "A file is mapped in multiples of the page size.  For a file
583		 * that is not a multiple of the page size, the remaining memory
584		 * is zeroed when mapped, and writes to that region are not
585		 * written out to the file."
586		 */
587		unsigned offset = i_size & (PAGE_SIZE - 1);
588
589		if (page->index > end_index || !offset)
590			goto confused;
591		zero_user_segment(page, offset, PAGE_SIZE);
592	}
593
594	/*
595	 * This page will go to BIO.  Do we need to send this BIO off first?
596	 */
597	if (bio && mpd->last_block_in_bio != blocks[0] - 1)
598		bio = mpage_bio_submit(wr, bio);
599
600alloc_new:
601	if (bio == NULL) {
602		if (first_unmapped == blocks_per_page) {
603			if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
604								page, wbc)) {
605				clean_buffers(page, first_unmapped);
606				goto out;
607			}
608		}
609		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
610				BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
611		if (bio == NULL)
612			goto confused;
613
614		wbc_init_bio(wbc, bio);
615	}
616
617	/*
618	 * Must try to add the page before marking the buffer clean or
619	 * the confused fail path above (OOM) will be very confused when
620	 * it finds all bh marked clean (i.e. it will not write anything)
621	 */
622	wbc_account_io(wbc, page, PAGE_SIZE);
623	length = first_unmapped << blkbits;
624	if (bio_add_page(bio, page, length, 0) < length) {
625		bio = mpage_bio_submit(wr, bio);
626		goto alloc_new;
627	}
628
629	clean_buffers(page, first_unmapped);
630
631	BUG_ON(PageWriteback(page));
632	set_page_writeback(page);
633	unlock_page(page);
634	if (boundary || (first_unmapped != blocks_per_page)) {
635		bio = mpage_bio_submit(wr, bio);
636		if (boundary_block) {
637			write_boundary_block(boundary_bdev,
638					boundary_block, 1 << blkbits);
639		}
640	} else {
641		mpd->last_block_in_bio = blocks[blocks_per_page - 1];
642	}
643	goto out;
644
645confused:
646	if (bio)
647		bio = mpage_bio_submit(wr, bio);
648
649	if (mpd->use_writepage) {
650		ret = mapping->a_ops->writepage(page, wbc);
651	} else {
652		ret = -EAGAIN;
653		goto out;
654	}
655	/*
656	 * The caller has a ref on the inode, so *mapping is stable
657	 */
658	mapping_set_error(mapping, ret);
659out:
660	mpd->bio = bio;
661	return ret;
662}
663
664/**
665 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
666 * @mapping: address space structure to write
667 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
668 * @get_block: the filesystem's block mapper function.
669 *             If this is NULL then use a_ops->writepage.  Otherwise, go
670 *             direct-to-BIO.
671 *
672 * This is a library function, which implements the writepages()
673 * address_space_operation.
674 *
675 * If a page is already under I/O, generic_writepages() skips it, even
676 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
677 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
678 * and msync() need to guarantee that all the data which was dirty at the time
679 * the call was made get new I/O started against them.  If wbc->sync_mode is
680 * WB_SYNC_ALL then we were called for data integrity and we must wait for
681 * existing IO to complete.
682 */
683int
684mpage_writepages(struct address_space *mapping,
685		struct writeback_control *wbc, get_block_t get_block)
686{
687	struct blk_plug plug;
688	int ret;
689
690	blk_start_plug(&plug);
691
692	if (!get_block)
693		ret = generic_writepages(mapping, wbc);
694	else {
695		struct mpage_data mpd = {
696			.bio = NULL,
697			.last_block_in_bio = 0,
698			.get_block = get_block,
699			.use_writepage = 1,
700		};
701
702		ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
703		if (mpd.bio) {
704			int wr = (wbc->sync_mode == WB_SYNC_ALL ?
705				  WRITE_SYNC : WRITE);
706			mpage_bio_submit(wr, mpd.bio);
707		}
708	}
709	blk_finish_plug(&plug);
710	return ret;
711}
712EXPORT_SYMBOL(mpage_writepages);
713
714int mpage_writepage(struct page *page, get_block_t get_block,
715	struct writeback_control *wbc)
716{
717	struct mpage_data mpd = {
718		.bio = NULL,
719		.last_block_in_bio = 0,
720		.get_block = get_block,
721		.use_writepage = 0,
722	};
723	int ret = __mpage_writepage(page, wbc, &mpd);
724	if (mpd.bio) {
725		int wr = (wbc->sync_mode == WB_SYNC_ALL ?
726			  WRITE_SYNC : WRITE);
727		mpage_bio_submit(wr, mpd.bio);
728	}
729	return ret;
730}
731EXPORT_SYMBOL(mpage_writepage);