Linux Audio

Check our new training course

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);
v5.9
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * fs/mpage.c
  4 *
  5 * Copyright (C) 2002, Linus Torvalds.
  6 *
  7 * Contains functions related to preparing and submitting BIOs which contain
  8 * multiple pagecache pages.
  9 *
 10 * 15May2002	Andrew Morton
 11 *		Initial version
 12 * 27Jun2002	axboe@suse.de
 13 *		use bio_add_page() to build bio's just the right size
 14 */
 15
 16#include <linux/kernel.h>
 17#include <linux/export.h>
 18#include <linux/mm.h>
 19#include <linux/kdev_t.h>
 20#include <linux/gfp.h>
 21#include <linux/bio.h>
 22#include <linux/fs.h>
 23#include <linux/buffer_head.h>
 24#include <linux/blkdev.h>
 25#include <linux/highmem.h>
 26#include <linux/prefetch.h>
 27#include <linux/mpage.h>
 28#include <linux/mm_inline.h>
 29#include <linux/writeback.h>
 30#include <linux/backing-dev.h>
 31#include <linux/pagevec.h>
 32#include <linux/cleancache.h>
 33#include "internal.h"
 34
 35/*
 36 * I/O completion handler for multipage BIOs.
 37 *
 38 * The mpage code never puts partial pages into a BIO (except for end-of-file).
 39 * If a page does not map to a contiguous run of blocks then it simply falls
 40 * back to block_read_full_page().
 41 *
 42 * Why is this?  If a page's completion depends on a number of different BIOs
 43 * which can complete in any order (or at the same time) then determining the
 44 * status of that page is hard.  See end_buffer_async_read() for the details.
 45 * There is no point in duplicating all that complexity.
 46 */
 47static void mpage_end_io(struct bio *bio)
 48{
 49	struct bio_vec *bv;
 50	struct bvec_iter_all iter_all;
 51
 52	bio_for_each_segment_all(bv, bio, iter_all) {
 53		struct page *page = bv->bv_page;
 54		page_endio(page, bio_op(bio),
 55			   blk_status_to_errno(bio->bi_status));
 56	}
 57
 58	bio_put(bio);
 59}
 60
 61static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
 62{
 63	bio->bi_end_io = mpage_end_io;
 64	bio_set_op_attrs(bio, op, op_flags);
 65	guard_bio_eod(bio);
 66	submit_bio(bio);
 67	return NULL;
 68}
 69
 70static struct bio *
 71mpage_alloc(struct block_device *bdev,
 72		sector_t first_sector, int nr_vecs,
 73		gfp_t gfp_flags)
 74{
 75	struct bio *bio;
 76
 77	/* Restrict the given (page cache) mask for slab allocations */
 78	gfp_flags &= GFP_KERNEL;
 79	bio = bio_alloc(gfp_flags, nr_vecs);
 80
 81	if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 82		while (!bio && (nr_vecs /= 2))
 83			bio = bio_alloc(gfp_flags, nr_vecs);
 84	}
 85
 86	if (bio) {
 87		bio_set_dev(bio, bdev);
 88		bio->bi_iter.bi_sector = first_sector;
 89	}
 90	return bio;
 91}
 92
 93/*
 94 * support function for mpage_readahead.  The fs supplied get_block might
 95 * return an up to date buffer.  This is used to map that buffer into
 96 * the page, which allows readpage to avoid triggering a duplicate call
 97 * to get_block.
 98 *
 99 * The idea is to avoid adding buffers to pages that don't already have
100 * them.  So when the buffer is up to date and the page size == block size,
101 * this marks the page up to date instead of adding new buffers.
102 */
103static void 
104map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block) 
105{
106	struct inode *inode = page->mapping->host;
107	struct buffer_head *page_bh, *head;
108	int block = 0;
109
110	if (!page_has_buffers(page)) {
111		/*
112		 * don't make any buffers if there is only one buffer on
113		 * the page and the page just needs to be set up to date
114		 */
115		if (inode->i_blkbits == PAGE_SHIFT &&
116		    buffer_uptodate(bh)) {
117			SetPageUptodate(page);    
118			return;
119		}
120		create_empty_buffers(page, i_blocksize(inode), 0);
121	}
122	head = page_buffers(page);
123	page_bh = head;
124	do {
125		if (block == page_block) {
126			page_bh->b_state = bh->b_state;
127			page_bh->b_bdev = bh->b_bdev;
128			page_bh->b_blocknr = bh->b_blocknr;
129			break;
130		}
131		page_bh = page_bh->b_this_page;
132		block++;
133	} while (page_bh != head);
134}
135
136struct mpage_readpage_args {
137	struct bio *bio;
138	struct page *page;
139	unsigned int nr_pages;
140	bool is_readahead;
141	sector_t last_block_in_bio;
142	struct buffer_head map_bh;
143	unsigned long first_logical_block;
144	get_block_t *get_block;
145};
146
147/*
148 * This is the worker routine which does all the work of mapping the disk
149 * blocks and constructs largest possible bios, submits them for IO if the
150 * blocks are not contiguous on the disk.
151 *
152 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
153 * represent the validity of its disk mapping and to decide when to do the next
154 * get_block() call.
155 */
156static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
 
 
 
 
157{
158	struct page *page = args->page;
159	struct inode *inode = page->mapping->host;
160	const unsigned blkbits = inode->i_blkbits;
161	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
162	const unsigned blocksize = 1 << blkbits;
163	struct buffer_head *map_bh = &args->map_bh;
164	sector_t block_in_file;
165	sector_t last_block;
166	sector_t last_block_in_file;
167	sector_t blocks[MAX_BUF_PER_PAGE];
168	unsigned page_block;
169	unsigned first_hole = blocks_per_page;
170	struct block_device *bdev = NULL;
171	int length;
172	int fully_mapped = 1;
173	int op_flags;
174	unsigned nblocks;
175	unsigned relative_block;
176	gfp_t gfp;
177
178	if (args->is_readahead) {
179		op_flags = REQ_RAHEAD;
180		gfp = readahead_gfp_mask(page->mapping);
181	} else {
182		op_flags = 0;
183		gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
184	}
185
186	if (page_has_buffers(page))
187		goto confused;
188
189	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
190	last_block = block_in_file + args->nr_pages * blocks_per_page;
191	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
192	if (last_block > last_block_in_file)
193		last_block = last_block_in_file;
194	page_block = 0;
195
196	/*
197	 * Map blocks using the result from the previous get_blocks call first.
198	 */
199	nblocks = map_bh->b_size >> blkbits;
200	if (buffer_mapped(map_bh) &&
201			block_in_file > args->first_logical_block &&
202			block_in_file < (args->first_logical_block + nblocks)) {
203		unsigned map_offset = block_in_file - args->first_logical_block;
204		unsigned last = nblocks - map_offset;
205
206		for (relative_block = 0; ; relative_block++) {
207			if (relative_block == last) {
208				clear_buffer_mapped(map_bh);
209				break;
210			}
211			if (page_block == blocks_per_page)
212				break;
213			blocks[page_block] = map_bh->b_blocknr + map_offset +
214						relative_block;
215			page_block++;
216			block_in_file++;
217		}
218		bdev = map_bh->b_bdev;
219	}
220
221	/*
222	 * Then do more get_blocks calls until we are done with this page.
223	 */
224	map_bh->b_page = page;
225	while (page_block < blocks_per_page) {
226		map_bh->b_state = 0;
227		map_bh->b_size = 0;
228
229		if (block_in_file < last_block) {
230			map_bh->b_size = (last_block-block_in_file) << blkbits;
231			if (args->get_block(inode, block_in_file, map_bh, 0))
232				goto confused;
233			args->first_logical_block = block_in_file;
234		}
235
236		if (!buffer_mapped(map_bh)) {
237			fully_mapped = 0;
238			if (first_hole == blocks_per_page)
239				first_hole = page_block;
240			page_block++;
241			block_in_file++;
242			continue;
243		}
244
245		/* some filesystems will copy data into the page during
246		 * the get_block call, in which case we don't want to
247		 * read it again.  map_buffer_to_page copies the data
248		 * we just collected from get_block into the page's buffers
249		 * so readpage doesn't have to repeat the get_block call
250		 */
251		if (buffer_uptodate(map_bh)) {
252			map_buffer_to_page(page, map_bh, page_block);
253			goto confused;
254		}
255	
256		if (first_hole != blocks_per_page)
257			goto confused;		/* hole -> non-hole */
258
259		/* Contiguous blocks? */
260		if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
261			goto confused;
262		nblocks = map_bh->b_size >> blkbits;
263		for (relative_block = 0; ; relative_block++) {
264			if (relative_block == nblocks) {
265				clear_buffer_mapped(map_bh);
266				break;
267			} else if (page_block == blocks_per_page)
268				break;
269			blocks[page_block] = map_bh->b_blocknr+relative_block;
270			page_block++;
271			block_in_file++;
272		}
273		bdev = map_bh->b_bdev;
274	}
275
276	if (first_hole != blocks_per_page) {
277		zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
278		if (first_hole == 0) {
279			SetPageUptodate(page);
280			unlock_page(page);
281			goto out;
282		}
283	} else if (fully_mapped) {
284		SetPageMappedToDisk(page);
285	}
286
287	if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
288	    cleancache_get_page(page) == 0) {
289		SetPageUptodate(page);
290		goto confused;
291	}
292
293	/*
294	 * This page will go to BIO.  Do we need to send this BIO off first?
295	 */
296	if (args->bio && (args->last_block_in_bio != blocks[0] - 1))
297		args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
298
299alloc_new:
300	if (args->bio == NULL) {
301		if (first_hole == blocks_per_page) {
302			if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
303								page))
304				goto out;
305		}
306		args->bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
307					min_t(int, args->nr_pages,
308					      BIO_MAX_PAGES),
309					gfp);
310		if (args->bio == NULL)
311			goto confused;
312	}
313
314	length = first_hole << blkbits;
315	if (bio_add_page(args->bio, page, length, 0) < length) {
316		args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
317		goto alloc_new;
318	}
319
320	relative_block = block_in_file - args->first_logical_block;
321	nblocks = map_bh->b_size >> blkbits;
322	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
323	    (first_hole != blocks_per_page))
324		args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
325	else
326		args->last_block_in_bio = blocks[blocks_per_page - 1];
327out:
328	return args->bio;
329
330confused:
331	if (args->bio)
332		args->bio = mpage_bio_submit(REQ_OP_READ, op_flags, args->bio);
333	if (!PageUptodate(page))
334		block_read_full_page(page, args->get_block);
335	else
336		unlock_page(page);
337	goto out;
338}
339
340/**
341 * mpage_readahead - start reads against pages
342 * @rac: Describes which pages to read.
 
 
 
 
 
343 * @get_block: The filesystem's block mapper function.
344 *
345 * This function walks the pages and the blocks within each page, building and
346 * emitting large BIOs.
347 *
348 * If anything unusual happens, such as:
349 *
350 * - encountering a page which has buffers
351 * - encountering a page which has a non-hole after a hole
352 * - encountering a page with non-contiguous blocks
353 *
354 * then this code just gives up and calls the buffer_head-based read function.
355 * It does handle a page which has holes at the end - that is a common case:
356 * the end-of-file on blocksize < PAGE_SIZE setups.
357 *
358 * BH_Boundary explanation:
359 *
360 * There is a problem.  The mpage read code assembles several pages, gets all
361 * their disk mappings, and then submits them all.  That's fine, but obtaining
362 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
363 *
364 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
365 * submitted in the following order:
366 *
367 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
368 *
369 * because the indirect block has to be read to get the mappings of blocks
370 * 13,14,15,16.  Obviously, this impacts performance.
371 *
372 * So what we do it to allow the filesystem's get_block() function to set
373 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
374 * after this one will require I/O against a block which is probably close to
375 * this one.  So you should push what I/O you have currently accumulated.
376 *
377 * This all causes the disk requests to be issued in the correct order.
378 */
379void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
 
 
380{
381	struct page *page;
382	struct mpage_readpage_args args = {
383		.get_block = get_block,
384		.is_readahead = true,
385	};
 
 
 
 
 
 
386
387	while ((page = readahead_page(rac))) {
388		prefetchw(&page->flags);
389		args.page = page;
390		args.nr_pages = readahead_count(rac);
391		args.bio = do_mpage_readpage(&args);
 
 
 
 
 
 
 
392		put_page(page);
393	}
394	if (args.bio)
395		mpage_bio_submit(REQ_OP_READ, REQ_RAHEAD, args.bio);
 
 
396}
397EXPORT_SYMBOL(mpage_readahead);
398
399/*
400 * This isn't called much at all
401 */
402int mpage_readpage(struct page *page, get_block_t get_block)
403{
404	struct mpage_readpage_args args = {
405		.page = page,
406		.nr_pages = 1,
407		.get_block = get_block,
408	};
409
410	args.bio = do_mpage_readpage(&args);
411	if (args.bio)
412		mpage_bio_submit(REQ_OP_READ, 0, args.bio);
 
 
 
413	return 0;
414}
415EXPORT_SYMBOL(mpage_readpage);
416
417/*
418 * Writing is not so simple.
419 *
420 * If the page has buffers then they will be used for obtaining the disk
421 * mapping.  We only support pages which are fully mapped-and-dirty, with a
422 * special case for pages which are unmapped at the end: end-of-file.
423 *
424 * If the page has no buffers (preferred) then the page is mapped here.
425 *
426 * If all blocks are found to be contiguous then the page can go into the
427 * BIO.  Otherwise fall back to the mapping's writepage().
428 * 
429 * FIXME: This code wants an estimate of how many pages are still to be
430 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
431 * just allocate full-size (16-page) BIOs.
432 */
433
434struct mpage_data {
435	struct bio *bio;
436	sector_t last_block_in_bio;
437	get_block_t *get_block;
438	unsigned use_writepage;
439};
440
441/*
442 * We have our BIO, so we can now mark the buffers clean.  Make
443 * sure to only clean buffers which we know we'll be writing.
444 */
445static void clean_buffers(struct page *page, unsigned first_unmapped)
446{
447	unsigned buffer_counter = 0;
448	struct buffer_head *bh, *head;
449	if (!page_has_buffers(page))
450		return;
451	head = page_buffers(page);
452	bh = head;
453
454	do {
455		if (buffer_counter++ == first_unmapped)
456			break;
457		clear_buffer_dirty(bh);
458		bh = bh->b_this_page;
459	} while (bh != head);
460
461	/*
462	 * we cannot drop the bh if the page is not uptodate or a concurrent
463	 * readpage would fail to serialize with the bh and it would read from
464	 * disk before we reach the platter.
465	 */
466	if (buffer_heads_over_limit && PageUptodate(page))
467		try_to_free_buffers(page);
468}
469
470/*
471 * For situations where we want to clean all buffers attached to a page.
472 * We don't need to calculate how many buffers are attached to the page,
473 * we just need to specify a number larger than the maximum number of buffers.
474 */
475void clean_page_buffers(struct page *page)
476{
477	clean_buffers(page, ~0U);
478}
479
480static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
481		      void *data)
482{
483	struct mpage_data *mpd = data;
484	struct bio *bio = mpd->bio;
485	struct address_space *mapping = page->mapping;
486	struct inode *inode = page->mapping->host;
487	const unsigned blkbits = inode->i_blkbits;
488	unsigned long end_index;
489	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
490	sector_t last_block;
491	sector_t block_in_file;
492	sector_t blocks[MAX_BUF_PER_PAGE];
493	unsigned page_block;
494	unsigned first_unmapped = blocks_per_page;
495	struct block_device *bdev = NULL;
496	int boundary = 0;
497	sector_t boundary_block = 0;
498	struct block_device *boundary_bdev = NULL;
499	int length;
500	struct buffer_head map_bh;
501	loff_t i_size = i_size_read(inode);
502	int ret = 0;
503	int op_flags = wbc_to_write_flags(wbc);
504
505	if (page_has_buffers(page)) {
506		struct buffer_head *head = page_buffers(page);
507		struct buffer_head *bh = head;
508
509		/* If they're all mapped and dirty, do it */
510		page_block = 0;
511		do {
512			BUG_ON(buffer_locked(bh));
513			if (!buffer_mapped(bh)) {
514				/*
515				 * unmapped dirty buffers are created by
516				 * __set_page_dirty_buffers -> mmapped data
517				 */
518				if (buffer_dirty(bh))
519					goto confused;
520				if (first_unmapped == blocks_per_page)
521					first_unmapped = page_block;
522				continue;
523			}
524
525			if (first_unmapped != blocks_per_page)
526				goto confused;	/* hole -> non-hole */
527
528			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
529				goto confused;
530			if (page_block) {
531				if (bh->b_blocknr != blocks[page_block-1] + 1)
532					goto confused;
533			}
534			blocks[page_block++] = bh->b_blocknr;
535			boundary = buffer_boundary(bh);
536			if (boundary) {
537				boundary_block = bh->b_blocknr;
538				boundary_bdev = bh->b_bdev;
539			}
540			bdev = bh->b_bdev;
541		} while ((bh = bh->b_this_page) != head);
542
543		if (first_unmapped)
544			goto page_is_mapped;
545
546		/*
547		 * Page has buffers, but they are all unmapped. The page was
548		 * created by pagein or read over a hole which was handled by
549		 * block_read_full_page().  If this address_space is also
550		 * using mpage_readahead then this can rarely happen.
551		 */
552		goto confused;
553	}
554
555	/*
556	 * The page has no buffers: map it to disk
557	 */
558	BUG_ON(!PageUptodate(page));
559	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
560	last_block = (i_size - 1) >> blkbits;
561	map_bh.b_page = page;
562	for (page_block = 0; page_block < blocks_per_page; ) {
563
564		map_bh.b_state = 0;
565		map_bh.b_size = 1 << blkbits;
566		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
567			goto confused;
568		if (buffer_new(&map_bh))
569			clean_bdev_bh_alias(&map_bh);
 
570		if (buffer_boundary(&map_bh)) {
571			boundary_block = map_bh.b_blocknr;
572			boundary_bdev = map_bh.b_bdev;
573		}
574		if (page_block) {
575			if (map_bh.b_blocknr != blocks[page_block-1] + 1)
576				goto confused;
577		}
578		blocks[page_block++] = map_bh.b_blocknr;
579		boundary = buffer_boundary(&map_bh);
580		bdev = map_bh.b_bdev;
581		if (block_in_file == last_block)
582			break;
583		block_in_file++;
584	}
585	BUG_ON(page_block == 0);
586
587	first_unmapped = page_block;
588
589page_is_mapped:
590	end_index = i_size >> PAGE_SHIFT;
591	if (page->index >= end_index) {
592		/*
593		 * The page straddles i_size.  It must be zeroed out on each
594		 * and every writepage invocation because it may be mmapped.
595		 * "A file is mapped in multiples of the page size.  For a file
596		 * that is not a multiple of the page size, the remaining memory
597		 * is zeroed when mapped, and writes to that region are not
598		 * written out to the file."
599		 */
600		unsigned offset = i_size & (PAGE_SIZE - 1);
601
602		if (page->index > end_index || !offset)
603			goto confused;
604		zero_user_segment(page, offset, PAGE_SIZE);
605	}
606
607	/*
608	 * This page will go to BIO.  Do we need to send this BIO off first?
609	 */
610	if (bio && mpd->last_block_in_bio != blocks[0] - 1)
611		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
612
613alloc_new:
614	if (bio == NULL) {
615		if (first_unmapped == blocks_per_page) {
616			if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
617								page, wbc))
 
618				goto out;
 
619		}
620		bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
621				BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
622		if (bio == NULL)
623			goto confused;
624
625		wbc_init_bio(wbc, bio);
626		bio->bi_write_hint = inode->i_write_hint;
627	}
628
629	/*
630	 * Must try to add the page before marking the buffer clean or
631	 * the confused fail path above (OOM) will be very confused when
632	 * it finds all bh marked clean (i.e. it will not write anything)
633	 */
634	wbc_account_cgroup_owner(wbc, page, PAGE_SIZE);
635	length = first_unmapped << blkbits;
636	if (bio_add_page(bio, page, length, 0) < length) {
637		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
638		goto alloc_new;
639	}
640
641	clean_buffers(page, first_unmapped);
642
643	BUG_ON(PageWriteback(page));
644	set_page_writeback(page);
645	unlock_page(page);
646	if (boundary || (first_unmapped != blocks_per_page)) {
647		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
648		if (boundary_block) {
649			write_boundary_block(boundary_bdev,
650					boundary_block, 1 << blkbits);
651		}
652	} else {
653		mpd->last_block_in_bio = blocks[blocks_per_page - 1];
654	}
655	goto out;
656
657confused:
658	if (bio)
659		bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
660
661	if (mpd->use_writepage) {
662		ret = mapping->a_ops->writepage(page, wbc);
663	} else {
664		ret = -EAGAIN;
665		goto out;
666	}
667	/*
668	 * The caller has a ref on the inode, so *mapping is stable
669	 */
670	mapping_set_error(mapping, ret);
671out:
672	mpd->bio = bio;
673	return ret;
674}
675
676/**
677 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
678 * @mapping: address space structure to write
679 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
680 * @get_block: the filesystem's block mapper function.
681 *             If this is NULL then use a_ops->writepage.  Otherwise, go
682 *             direct-to-BIO.
683 *
684 * This is a library function, which implements the writepages()
685 * address_space_operation.
686 *
687 * If a page is already under I/O, generic_writepages() skips it, even
688 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
689 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
690 * and msync() need to guarantee that all the data which was dirty at the time
691 * the call was made get new I/O started against them.  If wbc->sync_mode is
692 * WB_SYNC_ALL then we were called for data integrity and we must wait for
693 * existing IO to complete.
694 */
695int
696mpage_writepages(struct address_space *mapping,
697		struct writeback_control *wbc, get_block_t get_block)
698{
699	struct blk_plug plug;
700	int ret;
701
702	blk_start_plug(&plug);
703
704	if (!get_block)
705		ret = generic_writepages(mapping, wbc);
706	else {
707		struct mpage_data mpd = {
708			.bio = NULL,
709			.last_block_in_bio = 0,
710			.get_block = get_block,
711			.use_writepage = 1,
712		};
713
714		ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
715		if (mpd.bio) {
716			int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
717				  REQ_SYNC : 0);
718			mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
719		}
720	}
721	blk_finish_plug(&plug);
722	return ret;
723}
724EXPORT_SYMBOL(mpage_writepages);
725
726int mpage_writepage(struct page *page, get_block_t get_block,
727	struct writeback_control *wbc)
728{
729	struct mpage_data mpd = {
730		.bio = NULL,
731		.last_block_in_bio = 0,
732		.get_block = get_block,
733		.use_writepage = 0,
734	};
735	int ret = __mpage_writepage(page, wbc, &mpd);
736	if (mpd.bio) {
737		int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
738			  REQ_SYNC : 0);
739		mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
740	}
741	return ret;
742}
743EXPORT_SYMBOL(mpage_writepage);