Linux Audio

Check our new training course

Loading...
v6.2
  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 "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_folio().
 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	struct bvec_iter_all iter_all;
 50
 51	bio_for_each_segment_all(bv, bio, iter_all) {
 52		struct page *page = bv->bv_page;
 53		page_endio(page, bio_op(bio),
 54			   blk_status_to_errno(bio->bi_status));
 
 
 55	}
 56
 57	bio_put(bio);
 58}
 59
 60static struct bio *mpage_bio_submit(struct bio *bio)
 61{
 62	bio->bi_end_io = mpage_end_io;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 63	guard_bio_eod(bio);
 64	submit_bio(bio);
 65	return NULL;
 66}
 67
 68/*
 69 * support function for mpage_readahead.  The fs supplied get_block might
 70 * return an up to date buffer.  This is used to map that buffer into
 71 * the page, which allows read_folio to avoid triggering a duplicate call
 72 * to get_block.
 73 *
 74 * The idea is to avoid adding buffers to pages that don't already have
 75 * them.  So when the buffer is up to date and the page size == block size,
 76 * this marks the page up to date instead of adding new buffers.
 77 */
 78static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
 79		int page_block)
 80{
 81	struct inode *inode = folio->mapping->host;
 82	struct buffer_head *page_bh, *head;
 83	int block = 0;
 84
 85	head = folio_buffers(folio);
 86	if (!head) {
 87		/*
 88		 * don't make any buffers if there is only one buffer on
 89		 * the folio and the folio just needs to be set up to date
 90		 */
 91		if (inode->i_blkbits == PAGE_SHIFT &&
 92		    buffer_uptodate(bh)) {
 93			folio_mark_uptodate(folio);
 94			return;
 95		}
 96		create_empty_buffers(&folio->page, i_blocksize(inode), 0);
 97		head = folio_buffers(folio);
 98	}
 99
100	page_bh = head;
101	do {
102		if (block == page_block) {
103			page_bh->b_state = bh->b_state;
104			page_bh->b_bdev = bh->b_bdev;
105			page_bh->b_blocknr = bh->b_blocknr;
106			break;
107		}
108		page_bh = page_bh->b_this_page;
109		block++;
110	} while (page_bh != head);
111}
112
113struct mpage_readpage_args {
114	struct bio *bio;
115	struct folio *folio;
116	unsigned int nr_pages;
117	bool is_readahead;
118	sector_t last_block_in_bio;
119	struct buffer_head map_bh;
120	unsigned long first_logical_block;
121	get_block_t *get_block;
122};
123
124/*
125 * This is the worker routine which does all the work of mapping the disk
126 * blocks and constructs largest possible bios, submits them for IO if the
127 * blocks are not contiguous on the disk.
128 *
129 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
130 * represent the validity of its disk mapping and to decide when to do the next
131 * get_block() call.
132 */
133static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
134{
135	struct folio *folio = args->folio;
136	struct inode *inode = folio->mapping->host;
137	const unsigned blkbits = inode->i_blkbits;
138	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
139	const unsigned blocksize = 1 << blkbits;
140	struct buffer_head *map_bh = &args->map_bh;
141	sector_t block_in_file;
142	sector_t last_block;
143	sector_t last_block_in_file;
144	sector_t blocks[MAX_BUF_PER_PAGE];
145	unsigned page_block;
146	unsigned first_hole = blocks_per_page;
147	struct block_device *bdev = NULL;
148	int length;
149	int fully_mapped = 1;
150	blk_opf_t opf = REQ_OP_READ;
151	unsigned nblocks;
152	unsigned relative_block;
153	gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
154
155	/* MAX_BUF_PER_PAGE, for example */
156	VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
157
158	if (args->is_readahead) {
159		opf |= REQ_RAHEAD;
160		gfp |= __GFP_NORETRY | __GFP_NOWARN;
161	}
162
163	if (folio_buffers(folio))
164		goto confused;
165
166	block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
167	last_block = block_in_file + args->nr_pages * blocks_per_page;
168	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
169	if (last_block > last_block_in_file)
170		last_block = last_block_in_file;
171	page_block = 0;
172
173	/*
174	 * Map blocks using the result from the previous get_blocks call first.
175	 */
176	nblocks = map_bh->b_size >> blkbits;
177	if (buffer_mapped(map_bh) &&
178			block_in_file > args->first_logical_block &&
179			block_in_file < (args->first_logical_block + nblocks)) {
180		unsigned map_offset = block_in_file - args->first_logical_block;
181		unsigned last = nblocks - map_offset;
182
 
183		for (relative_block = 0; ; relative_block++) {
184			if (relative_block == last) {
185				clear_buffer_mapped(map_bh);
186				break;
187			}
188			if (page_block == blocks_per_page)
189				break;
190			blocks[page_block] = map_bh->b_blocknr + map_offset +
191						relative_block;
192			page_block++;
193			block_in_file++;
194		}
195		bdev = map_bh->b_bdev;
196	}
197
198	/*
199	 * Then do more get_blocks calls until we are done with this folio.
200	 */
201	map_bh->b_page = &folio->page;
202	while (page_block < blocks_per_page) {
203		map_bh->b_state = 0;
204		map_bh->b_size = 0;
205
206		if (block_in_file < last_block) {
207			map_bh->b_size = (last_block-block_in_file) << blkbits;
208			if (args->get_block(inode, block_in_file, map_bh, 0))
209				goto confused;
210			args->first_logical_block = block_in_file;
211		}
212
213		if (!buffer_mapped(map_bh)) {
214			fully_mapped = 0;
215			if (first_hole == blocks_per_page)
216				first_hole = page_block;
217			page_block++;
218			block_in_file++;
219			continue;
220		}
221
222		/* some filesystems will copy data into the page during
223		 * the get_block call, in which case we don't want to
224		 * read it again.  map_buffer_to_folio copies the data
225		 * we just collected from get_block into the folio's buffers
226		 * so read_folio doesn't have to repeat the get_block call
227		 */
228		if (buffer_uptodate(map_bh)) {
229			map_buffer_to_folio(folio, map_bh, page_block);
230			goto confused;
231		}
232	
233		if (first_hole != blocks_per_page)
234			goto confused;		/* hole -> non-hole */
235
236		/* Contiguous blocks? */
237		if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
 
 
238			goto confused;
239		nblocks = map_bh->b_size >> blkbits;
240		for (relative_block = 0; ; relative_block++) {
241			if (relative_block == nblocks) {
242				clear_buffer_mapped(map_bh);
243				break;
244			} else if (page_block == blocks_per_page)
245				break;
246			blocks[page_block] = map_bh->b_blocknr+relative_block;
247			page_block++;
248			block_in_file++;
249		}
250		bdev = map_bh->b_bdev;
251	}
252
253	if (first_hole != blocks_per_page) {
254		folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
255		if (first_hole == 0) {
256			folio_mark_uptodate(folio);
257			folio_unlock(folio);
258			goto out;
259		}
260	} else if (fully_mapped) {
261		folio_set_mappedtodisk(folio);
262	}
263
264	/*
265	 * This folio will go to BIO.  Do we need to send this BIO off first?
266	 */
267	if (args->bio && (args->last_block_in_bio != blocks[0] - 1))
268		args->bio = mpage_bio_submit(args->bio);
269
270alloc_new:
271	if (args->bio == NULL) {
272		if (first_hole == blocks_per_page) {
273			if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
274								&folio->page))
275				goto out;
276		}
277		args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
278				      gfp);
279		if (args->bio == NULL)
280			goto confused;
281		args->bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
282	}
283
284	length = first_hole << blkbits;
285	if (!bio_add_folio(args->bio, folio, length, 0)) {
286		args->bio = mpage_bio_submit(args->bio);
287		goto alloc_new;
288	}
289
290	relative_block = block_in_file - args->first_logical_block;
291	nblocks = map_bh->b_size >> blkbits;
292	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
293	    (first_hole != blocks_per_page))
294		args->bio = mpage_bio_submit(args->bio);
295	else
296		args->last_block_in_bio = blocks[blocks_per_page - 1];
297out:
298	return args->bio;
299
300confused:
301	if (args->bio)
302		args->bio = mpage_bio_submit(args->bio);
303	if (!folio_test_uptodate(folio))
304		block_read_full_folio(folio, args->get_block);
305	else
306		folio_unlock(folio);
307	goto out;
308}
309
310/**
311 * mpage_readahead - start reads against pages
312 * @rac: Describes which pages to read.
313 * @get_block: The filesystem's block mapper function.
314 *
315 * This function walks the pages and the blocks within each page, building and
316 * emitting large BIOs.
317 *
318 * If anything unusual happens, such as:
319 *
320 * - encountering a page which has buffers
321 * - encountering a page which has a non-hole after a hole
322 * - encountering a page with non-contiguous blocks
323 *
324 * then this code just gives up and calls the buffer_head-based read function.
325 * It does handle a page which has holes at the end - that is a common case:
326 * the end-of-file on blocksize < PAGE_SIZE setups.
327 *
328 * BH_Boundary explanation:
329 *
330 * There is a problem.  The mpage read code assembles several pages, gets all
331 * their disk mappings, and then submits them all.  That's fine, but obtaining
332 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
333 *
334 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
335 * submitted in the following order:
336 *
337 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
338 *
339 * because the indirect block has to be read to get the mappings of blocks
340 * 13,14,15,16.  Obviously, this impacts performance.
341 *
342 * So what we do it to allow the filesystem's get_block() function to set
343 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
344 * after this one will require I/O against a block which is probably close to
345 * this one.  So you should push what I/O you have currently accumulated.
346 *
347 * This all causes the disk requests to be issued in the correct order.
348 */
349void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
350{
351	struct folio *folio;
352	struct mpage_readpage_args args = {
353		.get_block = get_block,
354		.is_readahead = true,
355	};
356
357	while ((folio = readahead_folio(rac))) {
358		prefetchw(&folio->flags);
359		args.folio = folio;
360		args.nr_pages = readahead_count(rac);
361		args.bio = do_mpage_readpage(&args);
362	}
363	if (args.bio)
364		mpage_bio_submit(args.bio);
365}
366EXPORT_SYMBOL(mpage_readahead);
367
368/*
369 * This isn't called much at all
370 */
371int mpage_read_folio(struct folio *folio, get_block_t get_block)
372{
373	struct mpage_readpage_args args = {
374		.folio = folio,
375		.nr_pages = 1,
376		.get_block = get_block,
377	};
378
379	args.bio = do_mpage_readpage(&args);
380	if (args.bio)
381		mpage_bio_submit(args.bio);
382	return 0;
383}
384EXPORT_SYMBOL(mpage_read_folio);
385
386/*
387 * Writing is not so simple.
388 *
389 * If the page has buffers then they will be used for obtaining the disk
390 * mapping.  We only support pages which are fully mapped-and-dirty, with a
391 * special case for pages which are unmapped at the end: end-of-file.
392 *
393 * If the page has no buffers (preferred) then the page is mapped here.
394 *
395 * If all blocks are found to be contiguous then the page can go into the
396 * BIO.  Otherwise fall back to the mapping's writepage().
397 * 
398 * FIXME: This code wants an estimate of how many pages are still to be
399 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
400 * just allocate full-size (16-page) BIOs.
401 */
402
403struct mpage_data {
404	struct bio *bio;
405	sector_t last_block_in_bio;
406	get_block_t *get_block;
407};
408
409/*
410 * We have our BIO, so we can now mark the buffers clean.  Make
411 * sure to only clean buffers which we know we'll be writing.
412 */
413static void clean_buffers(struct page *page, unsigned first_unmapped)
414{
415	unsigned buffer_counter = 0;
416	struct buffer_head *bh, *head;
417	if (!page_has_buffers(page))
 
418		return;
419	head = page_buffers(page);
420	bh = head;
421
422	do {
423		if (buffer_counter++ == first_unmapped)
424			break;
425		clear_buffer_dirty(bh);
426		bh = bh->b_this_page;
427	} while (bh != head);
428
429	/*
430	 * we cannot drop the bh if the page is not uptodate or a concurrent
431	 * read_folio would fail to serialize with the bh and it would read from
432	 * disk before we reach the platter.
433	 */
434	if (buffer_heads_over_limit && PageUptodate(page))
435		try_to_free_buffers(page_folio(page));
436}
437
438/*
439 * For situations where we want to clean all buffers attached to a page.
440 * We don't need to calculate how many buffers are attached to the page,
441 * we just need to specify a number larger than the maximum number of buffers.
442 */
443void clean_page_buffers(struct page *page)
444{
445	clean_buffers(page, ~0U);
446}
447
448static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
449		      void *data)
450{
451	struct mpage_data *mpd = data;
452	struct bio *bio = mpd->bio;
453	struct address_space *mapping = page->mapping;
454	struct inode *inode = page->mapping->host;
455	const unsigned blkbits = inode->i_blkbits;
456	unsigned long end_index;
457	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
458	sector_t last_block;
459	sector_t block_in_file;
460	sector_t blocks[MAX_BUF_PER_PAGE];
461	unsigned page_block;
462	unsigned first_unmapped = blocks_per_page;
463	struct block_device *bdev = NULL;
464	int boundary = 0;
465	sector_t boundary_block = 0;
466	struct block_device *boundary_bdev = NULL;
467	int length;
468	struct buffer_head map_bh;
469	loff_t i_size = i_size_read(inode);
470	int ret = 0;
 
471
472	if (page_has_buffers(page)) {
473		struct buffer_head *head = page_buffers(page);
474		struct buffer_head *bh = head;
475
476		/* If they're all mapped and dirty, do it */
477		page_block = 0;
478		do {
479			BUG_ON(buffer_locked(bh));
480			if (!buffer_mapped(bh)) {
481				/*
482				 * unmapped dirty buffers are created by
483				 * block_dirty_folio -> mmapped data
484				 */
485				if (buffer_dirty(bh))
486					goto confused;
487				if (first_unmapped == blocks_per_page)
488					first_unmapped = page_block;
489				continue;
490			}
491
492			if (first_unmapped != blocks_per_page)
493				goto confused;	/* hole -> non-hole */
494
495			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
496				goto confused;
497			if (page_block) {
498				if (bh->b_blocknr != blocks[page_block-1] + 1)
499					goto confused;
 
 
500			}
501			blocks[page_block++] = bh->b_blocknr;
502			boundary = buffer_boundary(bh);
503			if (boundary) {
504				boundary_block = bh->b_blocknr;
505				boundary_bdev = bh->b_bdev;
506			}
507			bdev = bh->b_bdev;
508		} while ((bh = bh->b_this_page) != head);
509
510		if (first_unmapped)
511			goto page_is_mapped;
512
513		/*
514		 * Page has buffers, but they are all unmapped. The page was
515		 * created by pagein or read over a hole which was handled by
516		 * block_read_full_folio().  If this address_space is also
517		 * using mpage_readahead then this can rarely happen.
518		 */
519		goto confused;
520	}
521
522	/*
523	 * The page has no buffers: map it to disk
524	 */
525	BUG_ON(!PageUptodate(page));
526	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 
 
 
 
 
 
527	last_block = (i_size - 1) >> blkbits;
528	map_bh.b_page = page;
529	for (page_block = 0; page_block < blocks_per_page; ) {
530
531		map_bh.b_state = 0;
532		map_bh.b_size = 1 << blkbits;
533		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
534			goto confused;
 
 
535		if (buffer_new(&map_bh))
536			clean_bdev_bh_alias(&map_bh);
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_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_SIZE - 1);
568
569		if (page->index > end_index || !offset)
570			goto confused;
571		zero_user_segment(page, offset, PAGE_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(bio);
579
580alloc_new:
581	if (bio == NULL) {
582		if (first_unmapped == blocks_per_page) {
583			if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
584								page, wbc))
585				goto out;
586		}
587		bio = bio_alloc(bdev, BIO_MAX_VECS,
588				REQ_OP_WRITE | wbc_to_write_flags(wbc),
589				GFP_NOFS);
590		bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
591		wbc_init_bio(wbc, bio);
592	}
593
594	/*
595	 * Must try to add the page before marking the buffer clean or
596	 * the confused fail path above (OOM) will be very confused when
597	 * it finds all bh marked clean (i.e. it will not write anything)
598	 */
599	wbc_account_cgroup_owner(wbc, page, PAGE_SIZE);
600	length = first_unmapped << blkbits;
601	if (bio_add_page(bio, page, length, 0) < length) {
602		bio = mpage_bio_submit(bio);
603		goto alloc_new;
604	}
605
606	clean_buffers(page, first_unmapped);
607
608	BUG_ON(PageWriteback(page));
609	set_page_writeback(page);
610	unlock_page(page);
611	if (boundary || (first_unmapped != blocks_per_page)) {
612		bio = mpage_bio_submit(bio);
613		if (boundary_block) {
614			write_boundary_block(boundary_bdev,
615					boundary_block, 1 << blkbits);
616		}
617	} else {
618		mpd->last_block_in_bio = blocks[blocks_per_page - 1];
619	}
620	goto out;
621
622confused:
623	if (bio)
624		bio = mpage_bio_submit(bio);
625
626	/*
627	 * The caller has a ref on the inode, so *mapping is stable
628	 */
629	ret = block_write_full_page(page, mpd->get_block, wbc);
630	mapping_set_error(mapping, ret);
631out:
632	mpd->bio = bio;
633	return ret;
634}
635
636/**
637 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
638 * @mapping: address space structure to write
639 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
640 * @get_block: the filesystem's block mapper function.
641 *
642 * This is a library function, which implements the writepages()
643 * address_space_operation.
644 *
645 * If a page is already under I/O, generic_writepages() skips it, even
646 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
647 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
648 * and msync() need to guarantee that all the data which was dirty at the time
649 * the call was made get new I/O started against them.  If wbc->sync_mode is
650 * WB_SYNC_ALL then we were called for data integrity and we must wait for
651 * existing IO to complete.
652 */
653int
654mpage_writepages(struct address_space *mapping,
655		struct writeback_control *wbc, get_block_t get_block)
656{
657	struct mpage_data mpd = {
658		.get_block	= get_block,
659	};
660	struct blk_plug plug;
661	int ret;
662
663	blk_start_plug(&plug);
664	ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
665	if (mpd.bio)
666		mpage_bio_submit(mpd.bio);
667	blk_finish_plug(&plug);
668	return ret;
669}
670EXPORT_SYMBOL(mpage_writepages);
v6.8
  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 "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_folio().
 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_read_end_io(struct bio *bio)
 47{
 48	struct folio_iter fi;
 49	int err = blk_status_to_errno(bio->bi_status);
 50
 51	bio_for_each_folio_all(fi, bio) {
 52		if (err)
 53			folio_set_error(fi.folio);
 54		else
 55			folio_mark_uptodate(fi.folio);
 56		folio_unlock(fi.folio);
 57	}
 58
 59	bio_put(bio);
 60}
 61
 62static void mpage_write_end_io(struct bio *bio)
 63{
 64	struct folio_iter fi;
 65	int err = blk_status_to_errno(bio->bi_status);
 66
 67	bio_for_each_folio_all(fi, bio) {
 68		if (err) {
 69			folio_set_error(fi.folio);
 70			mapping_set_error(fi.folio->mapping, err);
 71		}
 72		folio_end_writeback(fi.folio);
 73	}
 74
 75	bio_put(bio);
 76}
 77
 78static struct bio *mpage_bio_submit_read(struct bio *bio)
 79{
 80	bio->bi_end_io = mpage_read_end_io;
 81	guard_bio_eod(bio);
 82	submit_bio(bio);
 83	return NULL;
 84}
 85
 86static struct bio *mpage_bio_submit_write(struct bio *bio)
 87{
 88	bio->bi_end_io = mpage_write_end_io;
 89	guard_bio_eod(bio);
 90	submit_bio(bio);
 91	return NULL;
 92}
 93
 94/*
 95 * support function for mpage_readahead.  The fs supplied get_block might
 96 * return an up to date buffer.  This is used to map that buffer into
 97 * the page, which allows read_folio to avoid triggering a duplicate call
 98 * to get_block.
 99 *
100 * The idea is to avoid adding buffers to pages that don't already have
101 * them.  So when the buffer is up to date and the page size == block size,
102 * this marks the page up to date instead of adding new buffers.
103 */
104static void map_buffer_to_folio(struct folio *folio, struct buffer_head *bh,
105		int page_block)
106{
107	struct inode *inode = folio->mapping->host;
108	struct buffer_head *page_bh, *head;
109	int block = 0;
110
111	head = folio_buffers(folio);
112	if (!head) {
113		/*
114		 * don't make any buffers if there is only one buffer on
115		 * the folio and the folio just needs to be set up to date
116		 */
117		if (inode->i_blkbits == PAGE_SHIFT &&
118		    buffer_uptodate(bh)) {
119			folio_mark_uptodate(folio);
120			return;
121		}
122		head = create_empty_buffers(folio, i_blocksize(inode), 0);
 
123	}
124
125	page_bh = head;
126	do {
127		if (block == page_block) {
128			page_bh->b_state = bh->b_state;
129			page_bh->b_bdev = bh->b_bdev;
130			page_bh->b_blocknr = bh->b_blocknr;
131			break;
132		}
133		page_bh = page_bh->b_this_page;
134		block++;
135	} while (page_bh != head);
136}
137
138struct mpage_readpage_args {
139	struct bio *bio;
140	struct folio *folio;
141	unsigned int nr_pages;
142	bool is_readahead;
143	sector_t last_block_in_bio;
144	struct buffer_head map_bh;
145	unsigned long first_logical_block;
146	get_block_t *get_block;
147};
148
149/*
150 * This is the worker routine which does all the work of mapping the disk
151 * blocks and constructs largest possible bios, submits them for IO if the
152 * blocks are not contiguous on the disk.
153 *
154 * We pass a buffer_head back and forth and use its buffer_mapped() flag to
155 * represent the validity of its disk mapping and to decide when to do the next
156 * get_block() call.
157 */
158static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
159{
160	struct folio *folio = args->folio;
161	struct inode *inode = folio->mapping->host;
162	const unsigned blkbits = inode->i_blkbits;
163	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
164	const unsigned blocksize = 1 << blkbits;
165	struct buffer_head *map_bh = &args->map_bh;
166	sector_t block_in_file;
167	sector_t last_block;
168	sector_t last_block_in_file;
169	sector_t first_block;
170	unsigned page_block;
171	unsigned first_hole = blocks_per_page;
172	struct block_device *bdev = NULL;
173	int length;
174	int fully_mapped = 1;
175	blk_opf_t opf = REQ_OP_READ;
176	unsigned nblocks;
177	unsigned relative_block;
178	gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
179
180	/* MAX_BUF_PER_PAGE, for example */
181	VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
182
183	if (args->is_readahead) {
184		opf |= REQ_RAHEAD;
185		gfp |= __GFP_NORETRY | __GFP_NOWARN;
186	}
187
188	if (folio_buffers(folio))
189		goto confused;
190
191	block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
192	last_block = block_in_file + args->nr_pages * blocks_per_page;
193	last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
194	if (last_block > last_block_in_file)
195		last_block = last_block_in_file;
196	page_block = 0;
197
198	/*
199	 * Map blocks using the result from the previous get_blocks call first.
200	 */
201	nblocks = map_bh->b_size >> blkbits;
202	if (buffer_mapped(map_bh) &&
203			block_in_file > args->first_logical_block &&
204			block_in_file < (args->first_logical_block + nblocks)) {
205		unsigned map_offset = block_in_file - args->first_logical_block;
206		unsigned last = nblocks - map_offset;
207
208		first_block = map_bh->b_blocknr + map_offset;
209		for (relative_block = 0; ; relative_block++) {
210			if (relative_block == last) {
211				clear_buffer_mapped(map_bh);
212				break;
213			}
214			if (page_block == blocks_per_page)
215				break;
 
 
216			page_block++;
217			block_in_file++;
218		}
219		bdev = map_bh->b_bdev;
220	}
221
222	/*
223	 * Then do more get_blocks calls until we are done with this folio.
224	 */
225	map_bh->b_folio = folio;
226	while (page_block < blocks_per_page) {
227		map_bh->b_state = 0;
228		map_bh->b_size = 0;
229
230		if (block_in_file < last_block) {
231			map_bh->b_size = (last_block-block_in_file) << blkbits;
232			if (args->get_block(inode, block_in_file, map_bh, 0))
233				goto confused;
234			args->first_logical_block = block_in_file;
235		}
236
237		if (!buffer_mapped(map_bh)) {
238			fully_mapped = 0;
239			if (first_hole == blocks_per_page)
240				first_hole = page_block;
241			page_block++;
242			block_in_file++;
243			continue;
244		}
245
246		/* some filesystems will copy data into the page during
247		 * the get_block call, in which case we don't want to
248		 * read it again.  map_buffer_to_folio copies the data
249		 * we just collected from get_block into the folio's buffers
250		 * so read_folio doesn't have to repeat the get_block call
251		 */
252		if (buffer_uptodate(map_bh)) {
253			map_buffer_to_folio(folio, map_bh, page_block);
254			goto confused;
255		}
256	
257		if (first_hole != blocks_per_page)
258			goto confused;		/* hole -> non-hole */
259
260		/* Contiguous blocks? */
261		if (!page_block)
262			first_block = map_bh->b_blocknr;
263		else if (first_block + page_block != map_bh->b_blocknr)
264			goto confused;
265		nblocks = map_bh->b_size >> blkbits;
266		for (relative_block = 0; ; relative_block++) {
267			if (relative_block == nblocks) {
268				clear_buffer_mapped(map_bh);
269				break;
270			} else if (page_block == blocks_per_page)
271				break;
 
272			page_block++;
273			block_in_file++;
274		}
275		bdev = map_bh->b_bdev;
276	}
277
278	if (first_hole != blocks_per_page) {
279		folio_zero_segment(folio, first_hole << blkbits, PAGE_SIZE);
280		if (first_hole == 0) {
281			folio_mark_uptodate(folio);
282			folio_unlock(folio);
283			goto out;
284		}
285	} else if (fully_mapped) {
286		folio_set_mappedtodisk(folio);
287	}
288
289	/*
290	 * This folio will go to BIO.  Do we need to send this BIO off first?
291	 */
292	if (args->bio && (args->last_block_in_bio != first_block - 1))
293		args->bio = mpage_bio_submit_read(args->bio);
294
295alloc_new:
296	if (args->bio == NULL) {
 
 
 
 
 
297		args->bio = bio_alloc(bdev, bio_max_segs(args->nr_pages), opf,
298				      gfp);
299		if (args->bio == NULL)
300			goto confused;
301		args->bio->bi_iter.bi_sector = first_block << (blkbits - 9);
302	}
303
304	length = first_hole << blkbits;
305	if (!bio_add_folio(args->bio, folio, length, 0)) {
306		args->bio = mpage_bio_submit_read(args->bio);
307		goto alloc_new;
308	}
309
310	relative_block = block_in_file - args->first_logical_block;
311	nblocks = map_bh->b_size >> blkbits;
312	if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
313	    (first_hole != blocks_per_page))
314		args->bio = mpage_bio_submit_read(args->bio);
315	else
316		args->last_block_in_bio = first_block + blocks_per_page - 1;
317out:
318	return args->bio;
319
320confused:
321	if (args->bio)
322		args->bio = mpage_bio_submit_read(args->bio);
323	if (!folio_test_uptodate(folio))
324		block_read_full_folio(folio, args->get_block);
325	else
326		folio_unlock(folio);
327	goto out;
328}
329
330/**
331 * mpage_readahead - start reads against pages
332 * @rac: Describes which pages to read.
333 * @get_block: The filesystem's block mapper function.
334 *
335 * This function walks the pages and the blocks within each page, building and
336 * emitting large BIOs.
337 *
338 * If anything unusual happens, such as:
339 *
340 * - encountering a page which has buffers
341 * - encountering a page which has a non-hole after a hole
342 * - encountering a page with non-contiguous blocks
343 *
344 * then this code just gives up and calls the buffer_head-based read function.
345 * It does handle a page which has holes at the end - that is a common case:
346 * the end-of-file on blocksize < PAGE_SIZE setups.
347 *
348 * BH_Boundary explanation:
349 *
350 * There is a problem.  The mpage read code assembles several pages, gets all
351 * their disk mappings, and then submits them all.  That's fine, but obtaining
352 * the disk mappings may require I/O.  Reads of indirect blocks, for example.
353 *
354 * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
355 * submitted in the following order:
356 *
357 * 	12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
358 *
359 * because the indirect block has to be read to get the mappings of blocks
360 * 13,14,15,16.  Obviously, this impacts performance.
361 *
362 * So what we do it to allow the filesystem's get_block() function to set
363 * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
364 * after this one will require I/O against a block which is probably close to
365 * this one.  So you should push what I/O you have currently accumulated.
366 *
367 * This all causes the disk requests to be issued in the correct order.
368 */
369void mpage_readahead(struct readahead_control *rac, get_block_t get_block)
370{
371	struct folio *folio;
372	struct mpage_readpage_args args = {
373		.get_block = get_block,
374		.is_readahead = true,
375	};
376
377	while ((folio = readahead_folio(rac))) {
378		prefetchw(&folio->flags);
379		args.folio = folio;
380		args.nr_pages = readahead_count(rac);
381		args.bio = do_mpage_readpage(&args);
382	}
383	if (args.bio)
384		mpage_bio_submit_read(args.bio);
385}
386EXPORT_SYMBOL(mpage_readahead);
387
388/*
389 * This isn't called much at all
390 */
391int mpage_read_folio(struct folio *folio, get_block_t get_block)
392{
393	struct mpage_readpage_args args = {
394		.folio = folio,
395		.nr_pages = 1,
396		.get_block = get_block,
397	};
398
399	args.bio = do_mpage_readpage(&args);
400	if (args.bio)
401		mpage_bio_submit_read(args.bio);
402	return 0;
403}
404EXPORT_SYMBOL(mpage_read_folio);
405
406/*
407 * Writing is not so simple.
408 *
409 * If the page has buffers then they will be used for obtaining the disk
410 * mapping.  We only support pages which are fully mapped-and-dirty, with a
411 * special case for pages which are unmapped at the end: end-of-file.
412 *
413 * If the page has no buffers (preferred) then the page is mapped here.
414 *
415 * If all blocks are found to be contiguous then the page can go into the
416 * BIO.  Otherwise fall back to the mapping's writepage().
417 * 
418 * FIXME: This code wants an estimate of how many pages are still to be
419 * written, so it can intelligently allocate a suitably-sized BIO.  For now,
420 * just allocate full-size (16-page) BIOs.
421 */
422
423struct mpage_data {
424	struct bio *bio;
425	sector_t last_block_in_bio;
426	get_block_t *get_block;
427};
428
429/*
430 * We have our BIO, so we can now mark the buffers clean.  Make
431 * sure to only clean buffers which we know we'll be writing.
432 */
433static void clean_buffers(struct folio *folio, unsigned first_unmapped)
434{
435	unsigned buffer_counter = 0;
436	struct buffer_head *bh, *head = folio_buffers(folio);
437
438	if (!head)
439		return;
 
440	bh = head;
441
442	do {
443		if (buffer_counter++ == first_unmapped)
444			break;
445		clear_buffer_dirty(bh);
446		bh = bh->b_this_page;
447	} while (bh != head);
448
449	/*
450	 * we cannot drop the bh if the page is not uptodate or a concurrent
451	 * read_folio would fail to serialize with the bh and it would read from
452	 * disk before we reach the platter.
453	 */
454	if (buffer_heads_over_limit && folio_test_uptodate(folio))
455		try_to_free_buffers(folio);
456}
457
458static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
 
 
 
 
 
 
 
 
 
 
459		      void *data)
460{
461	struct mpage_data *mpd = data;
462	struct bio *bio = mpd->bio;
463	struct address_space *mapping = folio->mapping;
464	struct inode *inode = mapping->host;
465	const unsigned blkbits = inode->i_blkbits;
 
466	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
467	sector_t last_block;
468	sector_t block_in_file;
469	sector_t first_block;
470	unsigned page_block;
471	unsigned first_unmapped = blocks_per_page;
472	struct block_device *bdev = NULL;
473	int boundary = 0;
474	sector_t boundary_block = 0;
475	struct block_device *boundary_bdev = NULL;
476	size_t length;
477	struct buffer_head map_bh;
478	loff_t i_size = i_size_read(inode);
479	int ret = 0;
480	struct buffer_head *head = folio_buffers(folio);
481
482	if (head) {
 
483		struct buffer_head *bh = head;
484
485		/* If they're all mapped and dirty, do it */
486		page_block = 0;
487		do {
488			BUG_ON(buffer_locked(bh));
489			if (!buffer_mapped(bh)) {
490				/*
491				 * unmapped dirty buffers are created by
492				 * block_dirty_folio -> mmapped data
493				 */
494				if (buffer_dirty(bh))
495					goto confused;
496				if (first_unmapped == blocks_per_page)
497					first_unmapped = page_block;
498				continue;
499			}
500
501			if (first_unmapped != blocks_per_page)
502				goto confused;	/* hole -> non-hole */
503
504			if (!buffer_dirty(bh) || !buffer_uptodate(bh))
505				goto confused;
506			if (page_block) {
507				if (bh->b_blocknr != first_block + page_block)
508					goto confused;
509			} else {
510				first_block = bh->b_blocknr;
511			}
512			page_block++;
513			boundary = buffer_boundary(bh);
514			if (boundary) {
515				boundary_block = bh->b_blocknr;
516				boundary_bdev = bh->b_bdev;
517			}
518			bdev = bh->b_bdev;
519		} while ((bh = bh->b_this_page) != head);
520
521		if (first_unmapped)
522			goto page_is_mapped;
523
524		/*
525		 * Page has buffers, but they are all unmapped. The page was
526		 * created by pagein or read over a hole which was handled by
527		 * block_read_full_folio().  If this address_space is also
528		 * using mpage_readahead then this can rarely happen.
529		 */
530		goto confused;
531	}
532
533	/*
534	 * The page has no buffers: map it to disk
535	 */
536	BUG_ON(!folio_test_uptodate(folio));
537	block_in_file = (sector_t)folio->index << (PAGE_SHIFT - blkbits);
538	/*
539	 * Whole page beyond EOF? Skip allocating blocks to avoid leaking
540	 * space.
541	 */
542	if (block_in_file >= (i_size + (1 << blkbits) - 1) >> blkbits)
543		goto page_is_mapped;
544	last_block = (i_size - 1) >> blkbits;
545	map_bh.b_folio = folio;
546	for (page_block = 0; page_block < blocks_per_page; ) {
547
548		map_bh.b_state = 0;
549		map_bh.b_size = 1 << blkbits;
550		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
551			goto confused;
552		if (!buffer_mapped(&map_bh))
553			goto confused;
554		if (buffer_new(&map_bh))
555			clean_bdev_bh_alias(&map_bh);
556		if (buffer_boundary(&map_bh)) {
557			boundary_block = map_bh.b_blocknr;
558			boundary_bdev = map_bh.b_bdev;
559		}
560		if (page_block) {
561			if (map_bh.b_blocknr != first_block + page_block)
562				goto confused;
563		} else {
564			first_block = map_bh.b_blocknr;
565		}
566		page_block++;
567		boundary = buffer_boundary(&map_bh);
568		bdev = map_bh.b_bdev;
569		if (block_in_file == last_block)
570			break;
571		block_in_file++;
572	}
573	BUG_ON(page_block == 0);
574
575	first_unmapped = page_block;
576
577page_is_mapped:
578	/* Don't bother writing beyond EOF, truncate will discard the folio */
579	if (folio_pos(folio) >= i_size)
580		goto confused;
581	length = folio_size(folio);
582	if (folio_pos(folio) + length > i_size) {
583		/*
584		 * The page straddles i_size.  It must be zeroed out on each
585		 * and every writepage invocation because it may be mmapped.
586		 * "A file is mapped in multiples of the page size.  For a file
587		 * that is not a multiple of the page size, the remaining memory
588		 * is zeroed when mapped, and writes to that region are not
589		 * written out to the file."
590		 */
591		length = i_size - folio_pos(folio);
592		folio_zero_segment(folio, length, folio_size(folio));
 
 
 
593	}
594
595	/*
596	 * This page will go to BIO.  Do we need to send this BIO off first?
597	 */
598	if (bio && mpd->last_block_in_bio != first_block - 1)
599		bio = mpage_bio_submit_write(bio);
600
601alloc_new:
602	if (bio == NULL) {
 
 
 
 
 
603		bio = bio_alloc(bdev, BIO_MAX_VECS,
604				REQ_OP_WRITE | wbc_to_write_flags(wbc),
605				GFP_NOFS);
606		bio->bi_iter.bi_sector = first_block << (blkbits - 9);
607		wbc_init_bio(wbc, bio);
608	}
609
610	/*
611	 * Must try to add the page before marking the buffer clean or
612	 * the confused fail path above (OOM) will be very confused when
613	 * it finds all bh marked clean (i.e. it will not write anything)
614	 */
615	wbc_account_cgroup_owner(wbc, &folio->page, folio_size(folio));
616	length = first_unmapped << blkbits;
617	if (!bio_add_folio(bio, folio, length, 0)) {
618		bio = mpage_bio_submit_write(bio);
619		goto alloc_new;
620	}
621
622	clean_buffers(folio, first_unmapped);
623
624	BUG_ON(folio_test_writeback(folio));
625	folio_start_writeback(folio);
626	folio_unlock(folio);
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 = first_block + blocks_per_page - 1;
635	}
636	goto out;
637
638confused:
639	if (bio)
640		bio = mpage_bio_submit_write(bio);
641
642	/*
643	 * The caller has a ref on the inode, so *mapping is stable
644	 */
645	ret = block_write_full_folio(folio, wbc, mpd->get_block);
646	mapping_set_error(mapping, ret);
647out:
648	mpd->bio = bio;
649	return ret;
650}
651
652/**
653 * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
654 * @mapping: address space structure to write
655 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
656 * @get_block: the filesystem's block mapper function.
657 *
658 * This is a library function, which implements the writepages()
659 * address_space_operation.
 
 
 
 
 
 
 
 
660 */
661int
662mpage_writepages(struct address_space *mapping,
663		struct writeback_control *wbc, get_block_t get_block)
664{
665	struct mpage_data mpd = {
666		.get_block	= get_block,
667	};
668	struct blk_plug plug;
669	int ret;
670
671	blk_start_plug(&plug);
672	ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
673	if (mpd.bio)
674		mpage_bio_submit_write(mpd.bio);
675	blk_finish_plug(&plug);
676	return ret;
677}
678EXPORT_SYMBOL(mpage_writepages);