Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Mar 24-27, 2025, special US time zones
Register
Loading...
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);
v3.1
  1/*
  2 * fs/mpage.c
  3 *
  4 * Copyright (C) 2002, Linus Torvalds.
  5 *
  6 * Contains functions related to preparing and submitting BIOs which contain
  7 * multiple pagecache pages.
  8 *
  9 * 15May2002	Andrew Morton
 10 *		Initial version
 11 * 27Jun2002	axboe@suse.de
 12 *		use bio_add_page() to build bio's just the right size
 13 */
 14
 15#include <linux/kernel.h>
 16#include <linux/module.h>
 17#include <linux/mm.h>
 18#include <linux/kdev_t.h>
 19#include <linux/gfp.h>
 20#include <linux/bio.h>
 21#include <linux/fs.h>
 22#include <linux/buffer_head.h>
 23#include <linux/blkdev.h>
 24#include <linux/highmem.h>
 25#include <linux/prefetch.h>
 26#include <linux/mpage.h>
 
 27#include <linux/writeback.h>
 28#include <linux/backing-dev.h>
 29#include <linux/pagevec.h>
 30#include <linux/cleancache.h>
 
 31
 32/*
 33 * I/O completion handler for multipage BIOs.
 34 *
 35 * The mpage code never puts partial pages into a BIO (except for end-of-file).
 36 * If a page does not map to a contiguous run of blocks then it simply falls
 37 * back to block_read_full_page().
 38 *
 39 * Why is this?  If a page's completion depends on a number of different BIOs
 40 * which can complete in any order (or at the same time) then determining the
 41 * status of that page is hard.  See end_buffer_async_read() for the details.
 42 * There is no point in duplicating all that complexity.
 43 */
 44static void mpage_end_io(struct bio *bio, int err)
 45{
 46	const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 47	struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
 48
 49	do {
 50		struct page *page = bvec->bv_page;
 
 
 51
 52		if (--bvec >= bio->bi_io_vec)
 53			prefetchw(&bvec->bv_page->flags);
 54		if (bio_data_dir(bio) == READ) {
 55			if (uptodate) {
 56				SetPageUptodate(page);
 57			} else {
 58				ClearPageUptodate(page);
 59				SetPageError(page);
 60			}
 61			unlock_page(page);
 62		} else { /* bio_data_dir(bio) == WRITE */
 63			if (!uptodate) {
 64				SetPageError(page);
 65				if (page->mapping)
 66					set_bit(AS_EIO, &page->mapping->flags);
 67			}
 68			end_page_writeback(page);
 69		}
 70	} while (bvec >= bio->bi_io_vec);
 71	bio_put(bio);
 72}
 73
 74static struct bio *mpage_bio_submit(int rw, struct bio *bio)
 75{
 76	bio->bi_end_io = mpage_end_io;
 
 77	submit_bio(rw, bio);
 78	return NULL;
 79}
 80
 81static struct bio *
 82mpage_alloc(struct block_device *bdev,
 83		sector_t first_sector, int nr_vecs,
 84		gfp_t gfp_flags)
 85{
 86	struct bio *bio;
 87
 88	bio = bio_alloc(gfp_flags, nr_vecs);
 89
 90	if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 91		while (!bio && (nr_vecs /= 2))
 92			bio = bio_alloc(gfp_flags, nr_vecs);
 93	}
 94
 95	if (bio) {
 96		bio->bi_bdev = bdev;
 97		bio->bi_sector = first_sector;
 98	}
 99	return bio;
100}
101
102/*
103 * support function for mpage_readpages.  The fs supplied get_block might
104 * return an up to date buffer.  This is used to map that buffer into
105 * the page, which allows readpage to avoid triggering a duplicate call
106 * to get_block.
107 *
108 * The idea is to avoid adding buffers to pages that don't already have
109 * them.  So when the buffer is up to date and the page size == block size,
110 * this marks the page up to date instead of adding new buffers.
111 */
112static void 
113map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
114{
115	struct inode *inode = page->mapping->host;
116	struct buffer_head *page_bh, *head;
117	int block = 0;
118
119	if (!page_has_buffers(page)) {
120		/*
121		 * don't make any buffers if there is only one buffer on
122		 * the page and the page just needs to be set up to date
123		 */
124		if (inode->i_blkbits == PAGE_CACHE_SHIFT && 
125		    buffer_uptodate(bh)) {
126			SetPageUptodate(page);    
127			return;
128		}
129		create_empty_buffers(page, 1 << inode->i_blkbits, 0);
130	}
131	head = page_buffers(page);
132	page_bh = head;
133	do {
134		if (block == page_block) {
135			page_bh->b_state = bh->b_state;
136			page_bh->b_bdev = bh->b_bdev;
137			page_bh->b_blocknr = bh->b_blocknr;
138			break;
139		}
140		page_bh = page_bh->b_this_page;
141		block++;
142	} while (page_bh != head);
143}
144
145/*
146 * This is the worker routine which does all the work of mapping the disk
147 * blocks and constructs largest possible bios, submits them for IO if the
148 * blocks are not contiguous on the disk.
149 *
150 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
151 * represent the validity of its disk mapping and to decide when to do the next
152 * get_block() call.
153 */
154static struct bio *
155do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
156		sector_t *last_block_in_bio, struct buffer_head *map_bh,
157		unsigned long *first_logical_block, get_block_t get_block)
 
158{
159	struct inode *inode = page->mapping->host;
160	const unsigned blkbits = inode->i_blkbits;
161	const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
162	const unsigned blocksize = 1 << blkbits;
163	sector_t block_in_file;
164	sector_t last_block;
165	sector_t last_block_in_file;
166	sector_t blocks[MAX_BUF_PER_PAGE];
167	unsigned page_block;
168	unsigned first_hole = blocks_per_page;
169	struct block_device *bdev = NULL;
170	int length;
171	int fully_mapped = 1;
172	unsigned nblocks;
173	unsigned relative_block;
174
175	if (page_has_buffers(page))
176		goto confused;
177
178	block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
179	last_block = block_in_file + nr_pages * blocks_per_page;
180	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
181	if (last_block > last_block_in_file)
182		last_block = last_block_in_file;
183	page_block = 0;
184
185	/*
186	 * Map blocks using the result from the previous get_blocks call first.
187	 */
188	nblocks = map_bh->b_size >> blkbits;
189	if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
190			block_in_file < (*first_logical_block + nblocks)) {
191		unsigned map_offset = block_in_file - *first_logical_block;
192		unsigned last = nblocks - map_offset;
193
194		for (relative_block = 0; ; relative_block++) {
195			if (relative_block == last) {
196				clear_buffer_mapped(map_bh);
197				break;
198			}
199			if (page_block == blocks_per_page)
200				break;
201			blocks[page_block] = map_bh->b_blocknr + map_offset +
202						relative_block;
203			page_block++;
204			block_in_file++;
205		}
206		bdev = map_bh->b_bdev;
207	}
208
209	/*
210	 * Then do more get_blocks calls until we are done with this page.
211	 */
212	map_bh->b_page = page;
213	while (page_block < blocks_per_page) {
214		map_bh->b_state = 0;
215		map_bh->b_size = 0;
216
217		if (block_in_file < last_block) {
218			map_bh->b_size = (last_block-block_in_file) << blkbits;
219			if (get_block(inode, block_in_file, map_bh, 0))
220				goto confused;
221			*first_logical_block = block_in_file;
222		}
223
224		if (!buffer_mapped(map_bh)) {
225			fully_mapped = 0;
226			if (first_hole == blocks_per_page)
227				first_hole = page_block;
228			page_block++;
229			block_in_file++;
230			continue;
231		}
232
233		/* some filesystems will copy data into the page during
234		 * the get_block call, in which case we don't want to
235		 * read it again.  map_buffer_to_page copies the data
236		 * we just collected from get_block into the page's buffers
237		 * so readpage doesn't have to repeat the get_block call
238		 */
239		if (buffer_uptodate(map_bh)) {
240			map_buffer_to_page(page, map_bh, page_block);
241			goto confused;
242		}
243	
244		if (first_hole != blocks_per_page)
245			goto confused;		/* hole -> non-hole */
246
247		/* Contiguous blocks? */
248		if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
249			goto confused;
250		nblocks = map_bh->b_size >> blkbits;
251		for (relative_block = 0; ; relative_block++) {
252			if (relative_block == nblocks) {
253				clear_buffer_mapped(map_bh);
254				break;
255			} else if (page_block == blocks_per_page)
256				break;
257			blocks[page_block] = map_bh->b_blocknr+relative_block;
258			page_block++;
259			block_in_file++;
260		}
261		bdev = map_bh->b_bdev;
262	}
263
264	if (first_hole != blocks_per_page) {
265		zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
266		if (first_hole == 0) {
267			SetPageUptodate(page);
268			unlock_page(page);
269			goto out;
270		}
271	} else if (fully_mapped) {
272		SetPageMappedToDisk(page);
273	}
274
275	if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
276	    cleancache_get_page(page) == 0) {
277		SetPageUptodate(page);
278		goto confused;
279	}
280
281	/*
282	 * This page will go to BIO.  Do we need to send this BIO off first?
283	 */
284	if (bio && (*last_block_in_bio != blocks[0] - 1))
285		bio = mpage_bio_submit(READ, bio);
286
287alloc_new:
288	if (bio == NULL) {
 
 
 
 
 
289		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
290			  	min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
291				GFP_KERNEL);
292		if (bio == NULL)
293			goto confused;
294	}
295
296	length = first_hole << blkbits;
297	if (bio_add_page(bio, page, length, 0) < length) {
298		bio = mpage_bio_submit(READ, bio);
299		goto alloc_new;
300	}
301
302	relative_block = block_in_file - *first_logical_block;
303	nblocks = map_bh->b_size >> blkbits;
304	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
305	    (first_hole != blocks_per_page))
306		bio = mpage_bio_submit(READ, bio);
307	else
308		*last_block_in_bio = blocks[blocks_per_page - 1];
309out:
310	return bio;
311
312confused:
313	if (bio)
314		bio = mpage_bio_submit(READ, bio);
315	if (!PageUptodate(page))
316	        block_read_full_page(page, get_block);
317	else
318		unlock_page(page);
319	goto out;
320}
321
322/**
323 * mpage_readpages - populate an address space with some pages & start reads against them
324 * @mapping: the address_space
325 * @pages: The address of a list_head which contains the target pages.  These
326 *   pages have their ->index populated and are otherwise uninitialised.
327 *   The page at @pages->prev has the lowest file offset, and reads should be
328 *   issued in @pages->prev to @pages->next order.
329 * @nr_pages: The number of pages at *@pages
330 * @get_block: The filesystem's block mapper function.
331 *
332 * This function walks the pages and the blocks within each page, building and
333 * emitting large BIOs.
334 *
335 * If anything unusual happens, such as:
336 *
337 * - encountering a page which has buffers
338 * - encountering a page which has a non-hole after a hole
339 * - encountering a page with non-contiguous blocks
340 *
341 * then this code just gives up and calls the buffer_head-based read function.
342 * It does handle a page which has holes at the end - that is a common case:
343 * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
344 *
345 * BH_Boundary explanation:
346 *
347 * There is a problem.  The mpage read code assembles several pages, gets all
348 * their disk mappings, and then submits them all.  That's fine, but obtaining
349 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
350 *
351 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
352 * submitted in the following order:
353 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
354 *
355 * because the indirect block has to be read to get the mappings of blocks
356 * 13,14,15,16.  Obviously, this impacts performance.
357 *
358 * So what we do it to allow the filesystem's get_block() function to set
359 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
360 * after this one will require I/O against a block which is probably close to
361 * this one.  So you should push what I/O you have currently accumulated.
362 *
363 * This all causes the disk requests to be issued in the correct order.
364 */
365int
366mpage_readpages(struct address_space *mapping, struct list_head *pages,
367				unsigned nr_pages, get_block_t get_block)
368{
369	struct bio *bio = NULL;
370	unsigned page_idx;
371	sector_t last_block_in_bio = 0;
372	struct buffer_head map_bh;
373	unsigned long first_logical_block = 0;
374	struct blk_plug plug;
375
376	blk_start_plug(&plug);
377
378	map_bh.b_state = 0;
379	map_bh.b_size = 0;
380	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
381		struct page *page = list_entry(pages->prev, struct page, lru);
382
383		prefetchw(&page->flags);
384		list_del(&page->lru);
385		if (!add_to_page_cache_lru(page, mapping,
386					page->index, GFP_KERNEL)) {
 
387			bio = do_mpage_readpage(bio, page,
388					nr_pages - page_idx,
389					&last_block_in_bio, &map_bh,
390					&first_logical_block,
391					get_block);
392		}
393		page_cache_release(page);
394	}
395	BUG_ON(!list_empty(pages));
396	if (bio)
397		mpage_bio_submit(READ, bio);
398	blk_finish_plug(&plug);
399	return 0;
400}
401EXPORT_SYMBOL(mpage_readpages);
402
403/*
404 * This isn't called much at all
405 */
406int mpage_readpage(struct page *page, get_block_t get_block)
407{
408	struct bio *bio = NULL;
409	sector_t last_block_in_bio = 0;
410	struct buffer_head map_bh;
411	unsigned long first_logical_block = 0;
 
412
413	map_bh.b_state = 0;
414	map_bh.b_size = 0;
415	bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
416			&map_bh, &first_logical_block, get_block);
417	if (bio)
418		mpage_bio_submit(READ, bio);
419	return 0;
420}
421EXPORT_SYMBOL(mpage_readpage);
422
423/*
424 * Writing is not so simple.
425 *
426 * If the page has buffers then they will be used for obtaining the disk
427 * mapping.  We only support pages which are fully mapped-and-dirty, with a
428 * special case for pages which are unmapped at the end: end-of-file.
429 *
430 * If the page has no buffers (preferred) then the page is mapped here.
431 *
432 * If all blocks are found to be contiguous then the page can go into the
433 * BIO.  Otherwise fall back to the mapping's writepage().
434 * 
435 * FIXME: This code wants an estimate of how many pages are still to be
436 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
437 * just allocate full-size (16-page) BIOs.
438 */
439
440struct mpage_data {
441	struct bio *bio;
442	sector_t last_block_in_bio;
443	get_block_t *get_block;
444	unsigned use_writepage;
445};
446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
447static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
448		      void *data)
449{
450	struct mpage_data *mpd = data;
451	struct bio *bio = mpd->bio;
452	struct address_space *mapping = page->mapping;
453	struct inode *inode = page->mapping->host;
454	const unsigned blkbits = inode->i_blkbits;
455	unsigned long end_index;
456	const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
457	sector_t last_block;
458	sector_t block_in_file;
459	sector_t blocks[MAX_BUF_PER_PAGE];
460	unsigned page_block;
461	unsigned first_unmapped = blocks_per_page;
462	struct block_device *bdev = NULL;
463	int boundary = 0;
464	sector_t boundary_block = 0;
465	struct block_device *boundary_bdev = NULL;
466	int length;
467	struct buffer_head map_bh;
468	loff_t i_size = i_size_read(inode);
469	int ret = 0;
 
470
471	if (page_has_buffers(page)) {
472		struct buffer_head *head = page_buffers(page);
473		struct buffer_head *bh = head;
474
475		/* If they're all mapped and dirty, do it */
476		page_block = 0;
477		do {
478			BUG_ON(buffer_locked(bh));
479			if (!buffer_mapped(bh)) {
480				/*
481				 * unmapped dirty buffers are created by
482				 * __set_page_dirty_buffers -> mmapped data
483				 */
484				if (buffer_dirty(bh))
485					goto confused;
486				if (first_unmapped == blocks_per_page)
487					first_unmapped = page_block;
488				continue;
489			}
490
491			if (first_unmapped != blocks_per_page)
492				goto confused;	/* hole -> non-hole */
493
494			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
495				goto confused;
496			if (page_block) {
497				if (bh->b_blocknr != blocks[page_block-1] + 1)
498					goto confused;
499			}
500			blocks[page_block++] = bh->b_blocknr;
501			boundary = buffer_boundary(bh);
502			if (boundary) {
503				boundary_block = bh->b_blocknr;
504				boundary_bdev = bh->b_bdev;
505			}
506			bdev = bh->b_bdev;
507		} while ((bh = bh->b_this_page) != head);
508
509		if (first_unmapped)
510			goto page_is_mapped;
511
512		/*
513		 * Page has buffers, but they are all unmapped. The page was
514		 * created by pagein or read over a hole which was handled by
515		 * block_read_full_page().  If this address_space is also
516		 * using mpage_readpages then this can rarely happen.
517		 */
518		goto confused;
519	}
520
521	/*
522	 * The page has no buffers: map it to disk
523	 */
524	BUG_ON(!PageUptodate(page));
525	block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
526	last_block = (i_size - 1) >> blkbits;
527	map_bh.b_page = page;
528	for (page_block = 0; page_block < blocks_per_page; ) {
529
530		map_bh.b_state = 0;
531		map_bh.b_size = 1 << blkbits;
532		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
533			goto confused;
534		if (buffer_new(&map_bh))
535			unmap_underlying_metadata(map_bh.b_bdev,
536						map_bh.b_blocknr);
537		if (buffer_boundary(&map_bh)) {
538			boundary_block = map_bh.b_blocknr;
539			boundary_bdev = map_bh.b_bdev;
540		}
541		if (page_block) {
542			if (map_bh.b_blocknr != blocks[page_block-1] + 1)
543				goto confused;
544		}
545		blocks[page_block++] = map_bh.b_blocknr;
546		boundary = buffer_boundary(&map_bh);
547		bdev = map_bh.b_bdev;
548		if (block_in_file == last_block)
549			break;
550		block_in_file++;
551	}
552	BUG_ON(page_block == 0);
553
554	first_unmapped = page_block;
555
556page_is_mapped:
557	end_index = i_size >> PAGE_CACHE_SHIFT;
558	if (page->index >= end_index) {
559		/*
560		 * The page straddles i_size.  It must be zeroed out on each
561		 * and every writepage invocation because it may be mmapped.
562		 * "A file is mapped in multiples of the page size.  For a file
563		 * that is not a multiple of the page size, the remaining memory
564		 * is zeroed when mapped, and writes to that region are not
565		 * written out to the file."
566		 */
567		unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
568
569		if (page->index > end_index || !offset)
570			goto confused;
571		zero_user_segment(page, offset, PAGE_CACHE_SIZE);
572	}
573
574	/*
575	 * This page will go to BIO.  Do we need to send this BIO off first?
576	 */
577	if (bio && mpd->last_block_in_bio != blocks[0] - 1)
578		bio = mpage_bio_submit(WRITE, bio);
579
580alloc_new:
581	if (bio == NULL) {
 
 
 
 
 
 
 
582		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
583				bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
584		if (bio == NULL)
585			goto confused;
 
 
586	}
587
588	/*
589	 * Must try to add the page before marking the buffer clean or
590	 * the confused fail path above (OOM) will be very confused when
591	 * it finds all bh marked clean (i.e. it will not write anything)
592	 */
 
593	length = first_unmapped << blkbits;
594	if (bio_add_page(bio, page, length, 0) < length) {
595		bio = mpage_bio_submit(WRITE, bio);
596		goto alloc_new;
597	}
598
599	/*
600	 * OK, we have our BIO, so we can now mark the buffers clean.  Make
601	 * sure to only clean buffers which we know we'll be writing.
602	 */
603	if (page_has_buffers(page)) {
604		struct buffer_head *head = page_buffers(page);
605		struct buffer_head *bh = head;
606		unsigned buffer_counter = 0;
607
608		do {
609			if (buffer_counter++ == first_unmapped)
610				break;
611			clear_buffer_dirty(bh);
612			bh = bh->b_this_page;
613		} while (bh != head);
614
615		/*
616		 * we cannot drop the bh if the page is not uptodate
617		 * or a concurrent readpage would fail to serialize with the bh
618		 * and it would read from disk before we reach the platter.
619		 */
620		if (buffer_heads_over_limit && PageUptodate(page))
621			try_to_free_buffers(page);
622	}
623
624	BUG_ON(PageWriteback(page));
625	set_page_writeback(page);
626	unlock_page(page);
627	if (boundary || (first_unmapped != blocks_per_page)) {
628		bio = mpage_bio_submit(WRITE, bio);
629		if (boundary_block) {
630			write_boundary_block(boundary_bdev,
631					boundary_block, 1 << blkbits);
632		}
633	} else {
634		mpd->last_block_in_bio = blocks[blocks_per_page - 1];
635	}
636	goto out;
637
638confused:
639	if (bio)
640		bio = mpage_bio_submit(WRITE, bio);
641
642	if (mpd->use_writepage) {
643		ret = mapping->a_ops->writepage(page, wbc);
644	} else {
645		ret = -EAGAIN;
646		goto out;
647	}
648	/*
649	 * The caller has a ref on the inode, so *mapping is stable
650	 */
651	mapping_set_error(mapping, ret);
652out:
653	mpd->bio = bio;
654	return ret;
655}
656
657/**
658 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
659 * @mapping: address space structure to write
660 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
661 * @get_block: the filesystem's block mapper function.
662 *             If this is NULL then use a_ops->writepage.  Otherwise, go
663 *             direct-to-BIO.
664 *
665 * This is a library function, which implements the writepages()
666 * address_space_operation.
667 *
668 * If a page is already under I/O, generic_writepages() skips it, even
669 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
670 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
671 * and msync() need to guarantee that all the data which was dirty at the time
672 * the call was made get new I/O started against them.  If wbc->sync_mode is
673 * WB_SYNC_ALL then we were called for data integrity and we must wait for
674 * existing IO to complete.
675 */
676int
677mpage_writepages(struct address_space *mapping,
678		struct writeback_control *wbc, get_block_t get_block)
679{
680	struct blk_plug plug;
681	int ret;
682
683	blk_start_plug(&plug);
684
685	if (!get_block)
686		ret = generic_writepages(mapping, wbc);
687	else {
688		struct mpage_data mpd = {
689			.bio = NULL,
690			.last_block_in_bio = 0,
691			.get_block = get_block,
692			.use_writepage = 1,
693		};
694
695		ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
696		if (mpd.bio)
697			mpage_bio_submit(WRITE, mpd.bio);
 
 
 
698	}
699	blk_finish_plug(&plug);
700	return ret;
701}
702EXPORT_SYMBOL(mpage_writepages);
703
704int mpage_writepage(struct page *page, get_block_t get_block,
705	struct writeback_control *wbc)
706{
707	struct mpage_data mpd = {
708		.bio = NULL,
709		.last_block_in_bio = 0,
710		.get_block = get_block,
711		.use_writepage = 0,
712	};
713	int ret = __mpage_writepage(page, wbc, &mpd);
714	if (mpd.bio)
715		mpage_bio_submit(WRITE, mpd.bio);
 
 
 
716	return ret;
717}
718EXPORT_SYMBOL(mpage_writepage);