buffered-io.c revision 740499c7
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (C) 2016-2019 Christoph Hellwig.
5 */
6#include <linux/module.h>
7#include <linux/compiler.h>
8#include <linux/fs.h>
9#include <linux/iomap.h>
10#include <linux/pagemap.h>
11#include <linux/uio.h>
12#include <linux/buffer_head.h>
13#include <linux/dax.h>
14#include <linux/writeback.h>
15#include <linux/list_sort.h>
16#include <linux/swap.h>
17#include <linux/bio.h>
18#include <linux/sched/signal.h>
19#include <linux/migrate.h>
20#include "trace.h"
21
22#include "../internal.h"
23
24/*
25 * Structure allocated for each page or THP when block size < page size
26 * to track sub-page uptodate status and I/O completions.
27 */
28struct iomap_page {
29	atomic_t		read_bytes_pending;
30	atomic_t		write_bytes_pending;
31	spinlock_t		uptodate_lock;
32	unsigned long		uptodate[];
33};
34
35static inline struct iomap_page *to_iomap_page(struct page *page)
36{
37	/*
38	 * per-block data is stored in the head page.  Callers should
39	 * not be dealing with tail pages, and if they are, they can
40	 * call thp_head() first.
41	 */
42	VM_BUG_ON_PGFLAGS(PageTail(page), page);
43
44	if (page_has_private(page))
45		return (struct iomap_page *)page_private(page);
46	return NULL;
47}
48
49static struct bio_set iomap_ioend_bioset;
50
51static struct iomap_page *
52iomap_page_create(struct inode *inode, struct page *page)
53{
54	struct iomap_page *iop = to_iomap_page(page);
55	unsigned int nr_blocks = i_blocks_per_page(inode, page);
56
57	if (iop || nr_blocks <= 1)
58		return iop;
59
60	iop = kzalloc(struct_size(iop, uptodate, BITS_TO_LONGS(nr_blocks)),
61			GFP_NOFS | __GFP_NOFAIL);
62	spin_lock_init(&iop->uptodate_lock);
63	if (PageUptodate(page))
64		bitmap_fill(iop->uptodate, nr_blocks);
65	attach_page_private(page, iop);
66	return iop;
67}
68
69static void
70iomap_page_release(struct page *page)
71{
72	struct iomap_page *iop = detach_page_private(page);
73	unsigned int nr_blocks = i_blocks_per_page(page->mapping->host, page);
74
75	if (!iop)
76		return;
77	WARN_ON_ONCE(atomic_read(&iop->read_bytes_pending));
78	WARN_ON_ONCE(atomic_read(&iop->write_bytes_pending));
79	WARN_ON_ONCE(bitmap_full(iop->uptodate, nr_blocks) !=
80			PageUptodate(page));
81	kfree(iop);
82}
83
84/*
85 * Calculate the range inside the page that we actually need to read.
86 */
87static void
88iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
89		loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
90{
91	loff_t orig_pos = *pos;
92	loff_t isize = i_size_read(inode);
93	unsigned block_bits = inode->i_blkbits;
94	unsigned block_size = (1 << block_bits);
95	unsigned poff = offset_in_page(*pos);
96	unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
97	unsigned first = poff >> block_bits;
98	unsigned last = (poff + plen - 1) >> block_bits;
99
100	/*
101	 * If the block size is smaller than the page size, we need to check the
102	 * per-block uptodate status and adjust the offset and length if needed
103	 * to avoid reading in already uptodate ranges.
104	 */
105	if (iop) {
106		unsigned int i;
107
108		/* move forward for each leading block marked uptodate */
109		for (i = first; i <= last; i++) {
110			if (!test_bit(i, iop->uptodate))
111				break;
112			*pos += block_size;
113			poff += block_size;
114			plen -= block_size;
115			first++;
116		}
117
118		/* truncate len if we find any trailing uptodate block(s) */
119		for ( ; i <= last; i++) {
120			if (test_bit(i, iop->uptodate)) {
121				plen -= (last - i + 1) * block_size;
122				last = i - 1;
123				break;
124			}
125		}
126	}
127
128	/*
129	 * If the extent spans the block that contains the i_size, we need to
130	 * handle both halves separately so that we properly zero data in the
131	 * page cache for blocks that are entirely outside of i_size.
132	 */
133	if (orig_pos <= isize && orig_pos + length > isize) {
134		unsigned end = offset_in_page(isize - 1) >> block_bits;
135
136		if (first <= end && last > end)
137			plen -= (last - end) * block_size;
138	}
139
140	*offp = poff;
141	*lenp = plen;
142}
143
144static void
145iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
146{
147	struct iomap_page *iop = to_iomap_page(page);
148	struct inode *inode = page->mapping->host;
149	unsigned first = off >> inode->i_blkbits;
150	unsigned last = (off + len - 1) >> inode->i_blkbits;
151	unsigned long flags;
152
153	spin_lock_irqsave(&iop->uptodate_lock, flags);
154	bitmap_set(iop->uptodate, first, last - first + 1);
155	if (bitmap_full(iop->uptodate, i_blocks_per_page(inode, page)))
156		SetPageUptodate(page);
157	spin_unlock_irqrestore(&iop->uptodate_lock, flags);
158}
159
160static void
161iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
162{
163	if (PageError(page))
164		return;
165
166	if (page_has_private(page))
167		iomap_iop_set_range_uptodate(page, off, len);
168	else
169		SetPageUptodate(page);
170}
171
172static void
173iomap_read_page_end_io(struct bio_vec *bvec, int error)
174{
175	struct page *page = bvec->bv_page;
176	struct iomap_page *iop = to_iomap_page(page);
177
178	if (unlikely(error)) {
179		ClearPageUptodate(page);
180		SetPageError(page);
181	} else {
182		iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
183	}
184
185	if (!iop || atomic_sub_and_test(bvec->bv_len, &iop->read_bytes_pending))
186		unlock_page(page);
187}
188
189static void
190iomap_read_end_io(struct bio *bio)
191{
192	int error = blk_status_to_errno(bio->bi_status);
193	struct bio_vec *bvec;
194	struct bvec_iter_all iter_all;
195
196	bio_for_each_segment_all(bvec, bio, iter_all)
197		iomap_read_page_end_io(bvec, error);
198	bio_put(bio);
199}
200
201struct iomap_readpage_ctx {
202	struct page		*cur_page;
203	bool			cur_page_in_bio;
204	struct bio		*bio;
205	struct readahead_control *rac;
206};
207
208static loff_t iomap_read_inline_data(struct inode *inode, struct page *page,
209		const struct iomap *iomap)
210{
211	size_t size = i_size_read(inode) - iomap->offset;
212	size_t poff = offset_in_page(iomap->offset);
213	void *addr;
214
215	if (PageUptodate(page))
216		return PAGE_SIZE - poff;
217
218	if (WARN_ON_ONCE(size > PAGE_SIZE - poff))
219		return -EIO;
220	if (WARN_ON_ONCE(size > PAGE_SIZE -
221			 offset_in_page(iomap->inline_data)))
222		return -EIO;
223	if (WARN_ON_ONCE(size > iomap->length))
224		return -EIO;
225	if (poff > 0)
226		iomap_page_create(inode, page);
227
228	addr = kmap_local_page(page) + poff;
229	memcpy(addr, iomap->inline_data, size);
230	memset(addr + size, 0, PAGE_SIZE - poff - size);
231	kunmap_local(addr);
232	iomap_set_range_uptodate(page, poff, PAGE_SIZE - poff);
233	return PAGE_SIZE - poff;
234}
235
236static inline bool iomap_block_needs_zeroing(struct inode *inode,
237		struct iomap *iomap, loff_t pos)
238{
239	return iomap->type != IOMAP_MAPPED ||
240		(iomap->flags & IOMAP_F_NEW) ||
241		pos >= i_size_read(inode);
242}
243
244static loff_t
245iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
246		struct iomap *iomap, struct iomap *srcmap)
247{
248	struct iomap_readpage_ctx *ctx = data;
249	struct page *page = ctx->cur_page;
250	struct iomap_page *iop;
251	loff_t orig_pos = pos;
252	unsigned poff, plen;
253	sector_t sector;
254
255	if (iomap->type == IOMAP_INLINE)
256		return min(iomap_read_inline_data(inode, page, iomap), length);
257
258	/* zero post-eof blocks as the page may be mapped */
259	iop = iomap_page_create(inode, page);
260	iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
261	if (plen == 0)
262		goto done;
263
264	if (iomap_block_needs_zeroing(inode, iomap, pos)) {
265		zero_user(page, poff, plen);
266		iomap_set_range_uptodate(page, poff, plen);
267		goto done;
268	}
269
270	ctx->cur_page_in_bio = true;
271	if (iop)
272		atomic_add(plen, &iop->read_bytes_pending);
273
274	sector = iomap_sector(iomap, pos);
275	if (!ctx->bio ||
276	    bio_end_sector(ctx->bio) != sector ||
277	    bio_add_page(ctx->bio, page, plen, poff) != plen) {
278		gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
279		gfp_t orig_gfp = gfp;
280		unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
281
282		if (ctx->bio)
283			submit_bio(ctx->bio);
284
285		if (ctx->rac) /* same as readahead_gfp_mask */
286			gfp |= __GFP_NORETRY | __GFP_NOWARN;
287		ctx->bio = bio_alloc(gfp, bio_max_segs(nr_vecs));
288		/*
289		 * If the bio_alloc fails, try it again for a single page to
290		 * avoid having to deal with partial page reads.  This emulates
291		 * what do_mpage_readpage does.
292		 */
293		if (!ctx->bio)
294			ctx->bio = bio_alloc(orig_gfp, 1);
295		ctx->bio->bi_opf = REQ_OP_READ;
296		if (ctx->rac)
297			ctx->bio->bi_opf |= REQ_RAHEAD;
298		ctx->bio->bi_iter.bi_sector = sector;
299		bio_set_dev(ctx->bio, iomap->bdev);
300		ctx->bio->bi_end_io = iomap_read_end_io;
301		__bio_add_page(ctx->bio, page, plen, poff);
302	}
303done:
304	/*
305	 * Move the caller beyond our range so that it keeps making progress.
306	 * For that, we have to include any leading non-uptodate ranges, but
307	 * we can skip trailing ones as they will be handled in the next
308	 * iteration.
309	 */
310	return pos - orig_pos + plen;
311}
312
313int
314iomap_readpage(struct page *page, const struct iomap_ops *ops)
315{
316	struct iomap_readpage_ctx ctx = { .cur_page = page };
317	struct inode *inode = page->mapping->host;
318	unsigned poff;
319	loff_t ret;
320
321	trace_iomap_readpage(page->mapping->host, 1);
322
323	for (poff = 0; poff < PAGE_SIZE; poff += ret) {
324		ret = iomap_apply(inode, page_offset(page) + poff,
325				PAGE_SIZE - poff, 0, ops, &ctx,
326				iomap_readpage_actor);
327		if (ret <= 0) {
328			WARN_ON_ONCE(ret == 0);
329			SetPageError(page);
330			break;
331		}
332	}
333
334	if (ctx.bio) {
335		submit_bio(ctx.bio);
336		WARN_ON_ONCE(!ctx.cur_page_in_bio);
337	} else {
338		WARN_ON_ONCE(ctx.cur_page_in_bio);
339		unlock_page(page);
340	}
341
342	/*
343	 * Just like mpage_readahead and block_read_full_page, we always
344	 * return 0 and just mark the page as PageError on errors.  This
345	 * should be cleaned up throughout the stack eventually.
346	 */
347	return 0;
348}
349EXPORT_SYMBOL_GPL(iomap_readpage);
350
351static loff_t
352iomap_readahead_actor(struct inode *inode, loff_t pos, loff_t length,
353		void *data, struct iomap *iomap, struct iomap *srcmap)
354{
355	struct iomap_readpage_ctx *ctx = data;
356	loff_t done, ret;
357
358	for (done = 0; done < length; done += ret) {
359		if (ctx->cur_page && offset_in_page(pos + done) == 0) {
360			if (!ctx->cur_page_in_bio)
361				unlock_page(ctx->cur_page);
362			put_page(ctx->cur_page);
363			ctx->cur_page = NULL;
364		}
365		if (!ctx->cur_page) {
366			ctx->cur_page = readahead_page(ctx->rac);
367			ctx->cur_page_in_bio = false;
368		}
369		ret = iomap_readpage_actor(inode, pos + done, length - done,
370				ctx, iomap, srcmap);
371	}
372
373	return done;
374}
375
376/**
377 * iomap_readahead - Attempt to read pages from a file.
378 * @rac: Describes the pages to be read.
379 * @ops: The operations vector for the filesystem.
380 *
381 * This function is for filesystems to call to implement their readahead
382 * address_space operation.
383 *
384 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
385 * blocks from disc), and may wait for it.  The caller may be trying to
386 * access a different page, and so sleeping excessively should be avoided.
387 * It may allocate memory, but should avoid costly allocations.  This
388 * function is called with memalloc_nofs set, so allocations will not cause
389 * the filesystem to be reentered.
390 */
391void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
392{
393	struct inode *inode = rac->mapping->host;
394	loff_t pos = readahead_pos(rac);
395	size_t length = readahead_length(rac);
396	struct iomap_readpage_ctx ctx = {
397		.rac	= rac,
398	};
399
400	trace_iomap_readahead(inode, readahead_count(rac));
401
402	while (length > 0) {
403		ssize_t ret = iomap_apply(inode, pos, length, 0, ops,
404				&ctx, iomap_readahead_actor);
405		if (ret <= 0) {
406			WARN_ON_ONCE(ret == 0);
407			break;
408		}
409		pos += ret;
410		length -= ret;
411	}
412
413	if (ctx.bio)
414		submit_bio(ctx.bio);
415	if (ctx.cur_page) {
416		if (!ctx.cur_page_in_bio)
417			unlock_page(ctx.cur_page);
418		put_page(ctx.cur_page);
419	}
420}
421EXPORT_SYMBOL_GPL(iomap_readahead);
422
423/*
424 * iomap_is_partially_uptodate checks whether blocks within a page are
425 * uptodate or not.
426 *
427 * Returns true if all blocks which correspond to a file portion
428 * we want to read within the page are uptodate.
429 */
430int
431iomap_is_partially_uptodate(struct page *page, unsigned long from,
432		unsigned long count)
433{
434	struct iomap_page *iop = to_iomap_page(page);
435	struct inode *inode = page->mapping->host;
436	unsigned len, first, last;
437	unsigned i;
438
439	/* Limit range to one page */
440	len = min_t(unsigned, PAGE_SIZE - from, count);
441
442	/* First and last blocks in range within page */
443	first = from >> inode->i_blkbits;
444	last = (from + len - 1) >> inode->i_blkbits;
445
446	if (iop) {
447		for (i = first; i <= last; i++)
448			if (!test_bit(i, iop->uptodate))
449				return 0;
450		return 1;
451	}
452
453	return 0;
454}
455EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
456
457int
458iomap_releasepage(struct page *page, gfp_t gfp_mask)
459{
460	trace_iomap_releasepage(page->mapping->host, page_offset(page),
461			PAGE_SIZE);
462
463	/*
464	 * mm accommodates an old ext3 case where clean pages might not have had
465	 * the dirty bit cleared. Thus, it can send actual dirty pages to
466	 * ->releasepage() via shrink_active_list(); skip those here.
467	 */
468	if (PageDirty(page) || PageWriteback(page))
469		return 0;
470	iomap_page_release(page);
471	return 1;
472}
473EXPORT_SYMBOL_GPL(iomap_releasepage);
474
475void
476iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
477{
478	trace_iomap_invalidatepage(page->mapping->host, offset, len);
479
480	/*
481	 * If we're invalidating the entire page, clear the dirty state from it
482	 * and release it to avoid unnecessary buildup of the LRU.
483	 */
484	if (offset == 0 && len == PAGE_SIZE) {
485		WARN_ON_ONCE(PageWriteback(page));
486		cancel_dirty_page(page);
487		iomap_page_release(page);
488	}
489}
490EXPORT_SYMBOL_GPL(iomap_invalidatepage);
491
492#ifdef CONFIG_MIGRATION
493int
494iomap_migrate_page(struct address_space *mapping, struct page *newpage,
495		struct page *page, enum migrate_mode mode)
496{
497	int ret;
498
499	ret = migrate_page_move_mapping(mapping, newpage, page, 0);
500	if (ret != MIGRATEPAGE_SUCCESS)
501		return ret;
502
503	if (page_has_private(page))
504		attach_page_private(newpage, detach_page_private(page));
505
506	if (mode != MIGRATE_SYNC_NO_COPY)
507		migrate_page_copy(newpage, page);
508	else
509		migrate_page_states(newpage, page);
510	return MIGRATEPAGE_SUCCESS;
511}
512EXPORT_SYMBOL_GPL(iomap_migrate_page);
513#endif /* CONFIG_MIGRATION */
514
515enum {
516	IOMAP_WRITE_F_UNSHARE		= (1 << 0),
517};
518
519static void
520iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
521{
522	loff_t i_size = i_size_read(inode);
523
524	/*
525	 * Only truncate newly allocated pages beyoned EOF, even if the
526	 * write started inside the existing inode size.
527	 */
528	if (pos + len > i_size)
529		truncate_pagecache_range(inode, max(pos, i_size), pos + len);
530}
531
532static int
533iomap_read_page_sync(loff_t block_start, struct page *page, unsigned poff,
534		unsigned plen, const struct iomap *iomap)
535{
536	struct bio_vec bvec;
537	struct bio bio;
538
539	bio_init(&bio, &bvec, 1);
540	bio.bi_opf = REQ_OP_READ;
541	bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
542	bio_set_dev(&bio, iomap->bdev);
543	__bio_add_page(&bio, page, plen, poff);
544	return submit_bio_wait(&bio);
545}
546
547static int
548__iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, int flags,
549		struct page *page, struct iomap *srcmap)
550{
551	struct iomap_page *iop = iomap_page_create(inode, page);
552	loff_t block_size = i_blocksize(inode);
553	loff_t block_start = round_down(pos, block_size);
554	loff_t block_end = round_up(pos + len, block_size);
555	unsigned from = offset_in_page(pos), to = from + len, poff, plen;
556
557	if (PageUptodate(page))
558		return 0;
559	ClearPageError(page);
560
561	do {
562		iomap_adjust_read_range(inode, iop, &block_start,
563				block_end - block_start, &poff, &plen);
564		if (plen == 0)
565			break;
566
567		if (!(flags & IOMAP_WRITE_F_UNSHARE) &&
568		    (from <= poff || from >= poff + plen) &&
569		    (to <= poff || to >= poff + plen))
570			continue;
571
572		if (iomap_block_needs_zeroing(inode, srcmap, block_start)) {
573			if (WARN_ON_ONCE(flags & IOMAP_WRITE_F_UNSHARE))
574				return -EIO;
575			zero_user_segments(page, poff, from, to, poff + plen);
576		} else {
577			int status = iomap_read_page_sync(block_start, page,
578					poff, plen, srcmap);
579			if (status)
580				return status;
581		}
582		iomap_set_range_uptodate(page, poff, plen);
583	} while ((block_start += plen) < block_end);
584
585	return 0;
586}
587
588static int iomap_write_begin_inline(struct inode *inode,
589		struct page *page, struct iomap *srcmap)
590{
591	int ret;
592
593	/* needs more work for the tailpacking case; disable for now */
594	if (WARN_ON_ONCE(srcmap->offset != 0))
595		return -EIO;
596	ret = iomap_read_inline_data(inode, page, srcmap);
597	if (ret < 0)
598		return ret;
599	return 0;
600}
601
602static int
603iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
604		struct page **pagep, struct iomap *iomap, struct iomap *srcmap)
605{
606	const struct iomap_page_ops *page_ops = iomap->page_ops;
607	struct page *page;
608	int status = 0;
609
610	BUG_ON(pos + len > iomap->offset + iomap->length);
611	if (srcmap != iomap)
612		BUG_ON(pos + len > srcmap->offset + srcmap->length);
613
614	if (fatal_signal_pending(current))
615		return -EINTR;
616
617	if (page_ops && page_ops->page_prepare) {
618		status = page_ops->page_prepare(inode, pos, len);
619		if (status)
620			return status;
621	}
622
623	page = grab_cache_page_write_begin(inode->i_mapping, pos >> PAGE_SHIFT,
624			AOP_FLAG_NOFS);
625	if (!page) {
626		status = -ENOMEM;
627		goto out_no_page;
628	}
629
630	if (srcmap->type == IOMAP_INLINE)
631		status = iomap_write_begin_inline(inode, page, srcmap);
632	else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
633		status = __block_write_begin_int(page, pos, len, NULL, srcmap);
634	else
635		status = __iomap_write_begin(inode, pos, len, flags, page,
636				srcmap);
637
638	if (unlikely(status))
639		goto out_unlock;
640
641	*pagep = page;
642	return 0;
643
644out_unlock:
645	unlock_page(page);
646	put_page(page);
647	iomap_write_failed(inode, pos, len);
648
649out_no_page:
650	if (page_ops && page_ops->page_done)
651		page_ops->page_done(inode, pos, 0, NULL);
652	return status;
653}
654
655static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
656		size_t copied, struct page *page)
657{
658	flush_dcache_page(page);
659
660	/*
661	 * The blocks that were entirely written will now be uptodate, so we
662	 * don't have to worry about a readpage reading them and overwriting a
663	 * partial write.  However, if we've encountered a short write and only
664	 * partially written into a block, it will not be marked uptodate, so a
665	 * readpage might come in and destroy our partial write.
666	 *
667	 * Do the simplest thing and just treat any short write to a
668	 * non-uptodate page as a zero-length write, and force the caller to
669	 * redo the whole thing.
670	 */
671	if (unlikely(copied < len && !PageUptodate(page)))
672		return 0;
673	iomap_set_range_uptodate(page, offset_in_page(pos), len);
674	__set_page_dirty_nobuffers(page);
675	return copied;
676}
677
678static size_t iomap_write_end_inline(struct inode *inode, struct page *page,
679		struct iomap *iomap, loff_t pos, size_t copied)
680{
681	void *addr;
682
683	WARN_ON_ONCE(!PageUptodate(page));
684	BUG_ON(!iomap_inline_data_valid(iomap));
685
686	flush_dcache_page(page);
687	addr = kmap_local_page(page) + pos;
688	memcpy(iomap_inline_data(iomap, pos), addr, copied);
689	kunmap_local(addr);
690
691	mark_inode_dirty(inode);
692	return copied;
693}
694
695/* Returns the number of bytes copied.  May be 0.  Cannot be an errno. */
696static size_t iomap_write_end(struct inode *inode, loff_t pos, size_t len,
697		size_t copied, struct page *page, struct iomap *iomap,
698		struct iomap *srcmap)
699{
700	const struct iomap_page_ops *page_ops = iomap->page_ops;
701	loff_t old_size = inode->i_size;
702	size_t ret;
703
704	if (srcmap->type == IOMAP_INLINE) {
705		ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
706	} else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
707		ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
708				page, NULL);
709	} else {
710		ret = __iomap_write_end(inode, pos, len, copied, page);
711	}
712
713	/*
714	 * Update the in-memory inode size after copying the data into the page
715	 * cache.  It's up to the file system to write the updated size to disk,
716	 * preferably after I/O completion so that no stale data is exposed.
717	 */
718	if (pos + ret > old_size) {
719		i_size_write(inode, pos + ret);
720		iomap->flags |= IOMAP_F_SIZE_CHANGED;
721	}
722	unlock_page(page);
723
724	if (old_size < pos)
725		pagecache_isize_extended(inode, old_size, pos);
726	if (page_ops && page_ops->page_done)
727		page_ops->page_done(inode, pos, ret, page);
728	put_page(page);
729
730	if (ret < len)
731		iomap_write_failed(inode, pos, len);
732	return ret;
733}
734
735static loff_t
736iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
737		struct iomap *iomap, struct iomap *srcmap)
738{
739	struct iov_iter *i = data;
740	long status = 0;
741	ssize_t written = 0;
742
743	do {
744		struct page *page;
745		unsigned long offset;	/* Offset into pagecache page */
746		unsigned long bytes;	/* Bytes to write to page */
747		size_t copied;		/* Bytes copied from user */
748
749		offset = offset_in_page(pos);
750		bytes = min_t(unsigned long, PAGE_SIZE - offset,
751						iov_iter_count(i));
752again:
753		if (bytes > length)
754			bytes = length;
755
756		/*
757		 * Bring in the user page that we'll copy from _first_.
758		 * Otherwise there's a nasty deadlock on copying from the
759		 * same page as we're writing to, without it being marked
760		 * up-to-date.
761		 */
762		if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
763			status = -EFAULT;
764			break;
765		}
766
767		status = iomap_write_begin(inode, pos, bytes, 0, &page, iomap,
768				srcmap);
769		if (unlikely(status))
770			break;
771
772		if (mapping_writably_mapped(inode->i_mapping))
773			flush_dcache_page(page);
774
775		copied = copy_page_from_iter_atomic(page, offset, bytes, i);
776
777		status = iomap_write_end(inode, pos, bytes, copied, page, iomap,
778				srcmap);
779
780		if (unlikely(copied != status))
781			iov_iter_revert(i, copied - status);
782
783		cond_resched();
784		if (unlikely(status == 0)) {
785			/*
786			 * A short copy made iomap_write_end() reject the
787			 * thing entirely.  Might be memory poisoning
788			 * halfway through, might be a race with munmap,
789			 * might be severe memory pressure.
790			 */
791			if (copied)
792				bytes = copied;
793			goto again;
794		}
795		pos += status;
796		written += status;
797		length -= status;
798
799		balance_dirty_pages_ratelimited(inode->i_mapping);
800	} while (iov_iter_count(i) && length);
801
802	return written ? written : status;
803}
804
805ssize_t
806iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
807		const struct iomap_ops *ops)
808{
809	struct inode *inode = iocb->ki_filp->f_mapping->host;
810	loff_t pos = iocb->ki_pos, ret = 0, written = 0;
811
812	while (iov_iter_count(iter)) {
813		ret = iomap_apply(inode, pos, iov_iter_count(iter),
814				IOMAP_WRITE, ops, iter, iomap_write_actor);
815		if (ret <= 0)
816			break;
817		pos += ret;
818		written += ret;
819	}
820
821	return written ? written : ret;
822}
823EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
824
825static loff_t
826iomap_unshare_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
827		struct iomap *iomap, struct iomap *srcmap)
828{
829	long status = 0;
830	loff_t written = 0;
831
832	/* don't bother with blocks that are not shared to start with */
833	if (!(iomap->flags & IOMAP_F_SHARED))
834		return length;
835	/* don't bother with holes or unwritten extents */
836	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
837		return length;
838
839	do {
840		unsigned long offset = offset_in_page(pos);
841		unsigned long bytes = min_t(loff_t, PAGE_SIZE - offset, length);
842		struct page *page;
843
844		status = iomap_write_begin(inode, pos, bytes,
845				IOMAP_WRITE_F_UNSHARE, &page, iomap, srcmap);
846		if (unlikely(status))
847			return status;
848
849		status = iomap_write_end(inode, pos, bytes, bytes, page, iomap,
850				srcmap);
851		if (WARN_ON_ONCE(status == 0))
852			return -EIO;
853
854		cond_resched();
855
856		pos += status;
857		written += status;
858		length -= status;
859
860		balance_dirty_pages_ratelimited(inode->i_mapping);
861	} while (length);
862
863	return written;
864}
865
866int
867iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
868		const struct iomap_ops *ops)
869{
870	loff_t ret;
871
872	while (len) {
873		ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
874				iomap_unshare_actor);
875		if (ret <= 0)
876			return ret;
877		pos += ret;
878		len -= ret;
879	}
880
881	return 0;
882}
883EXPORT_SYMBOL_GPL(iomap_file_unshare);
884
885static s64 iomap_zero(struct inode *inode, loff_t pos, u64 length,
886		struct iomap *iomap, struct iomap *srcmap)
887{
888	struct page *page;
889	int status;
890	unsigned offset = offset_in_page(pos);
891	unsigned bytes = min_t(u64, PAGE_SIZE - offset, length);
892
893	status = iomap_write_begin(inode, pos, bytes, 0, &page, iomap, srcmap);
894	if (status)
895		return status;
896
897	zero_user(page, offset, bytes);
898	mark_page_accessed(page);
899
900	return iomap_write_end(inode, pos, bytes, bytes, page, iomap, srcmap);
901}
902
903static loff_t iomap_zero_range_actor(struct inode *inode, loff_t pos,
904		loff_t length, void *data, struct iomap *iomap,
905		struct iomap *srcmap)
906{
907	bool *did_zero = data;
908	loff_t written = 0;
909
910	/* already zeroed?  we're done. */
911	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
912		return length;
913
914	do {
915		s64 bytes;
916
917		if (IS_DAX(inode))
918			bytes = dax_iomap_zero(pos, length, iomap);
919		else
920			bytes = iomap_zero(inode, pos, length, iomap, srcmap);
921		if (bytes < 0)
922			return bytes;
923
924		pos += bytes;
925		length -= bytes;
926		written += bytes;
927		if (did_zero)
928			*did_zero = true;
929	} while (length > 0);
930
931	return written;
932}
933
934int
935iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
936		const struct iomap_ops *ops)
937{
938	loff_t ret;
939
940	while (len > 0) {
941		ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
942				ops, did_zero, iomap_zero_range_actor);
943		if (ret <= 0)
944			return ret;
945
946		pos += ret;
947		len -= ret;
948	}
949
950	return 0;
951}
952EXPORT_SYMBOL_GPL(iomap_zero_range);
953
954int
955iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
956		const struct iomap_ops *ops)
957{
958	unsigned int blocksize = i_blocksize(inode);
959	unsigned int off = pos & (blocksize - 1);
960
961	/* Block boundary? Nothing to do */
962	if (!off)
963		return 0;
964	return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
965}
966EXPORT_SYMBOL_GPL(iomap_truncate_page);
967
968static loff_t
969iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
970		void *data, struct iomap *iomap, struct iomap *srcmap)
971{
972	struct page *page = data;
973	int ret;
974
975	if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
976		ret = __block_write_begin_int(page, pos, length, NULL, iomap);
977		if (ret)
978			return ret;
979		block_commit_write(page, 0, length);
980	} else {
981		WARN_ON_ONCE(!PageUptodate(page));
982		set_page_dirty(page);
983	}
984
985	return length;
986}
987
988vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
989{
990	struct page *page = vmf->page;
991	struct inode *inode = file_inode(vmf->vma->vm_file);
992	unsigned long length;
993	loff_t offset;
994	ssize_t ret;
995
996	lock_page(page);
997	ret = page_mkwrite_check_truncate(page, inode);
998	if (ret < 0)
999		goto out_unlock;
1000	length = ret;
1001
1002	offset = page_offset(page);
1003	while (length > 0) {
1004		ret = iomap_apply(inode, offset, length,
1005				IOMAP_WRITE | IOMAP_FAULT, ops, page,
1006				iomap_page_mkwrite_actor);
1007		if (unlikely(ret <= 0))
1008			goto out_unlock;
1009		offset += ret;
1010		length -= ret;
1011	}
1012
1013	wait_for_stable_page(page);
1014	return VM_FAULT_LOCKED;
1015out_unlock:
1016	unlock_page(page);
1017	return block_page_mkwrite_return(ret);
1018}
1019EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1020
1021static void
1022iomap_finish_page_writeback(struct inode *inode, struct page *page,
1023		int error, unsigned int len)
1024{
1025	struct iomap_page *iop = to_iomap_page(page);
1026
1027	if (error) {
1028		SetPageError(page);
1029		mapping_set_error(inode->i_mapping, error);
1030	}
1031
1032	WARN_ON_ONCE(i_blocks_per_page(inode, page) > 1 && !iop);
1033	WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) <= 0);
1034
1035	if (!iop || atomic_sub_and_test(len, &iop->write_bytes_pending))
1036		end_page_writeback(page);
1037}
1038
1039/*
1040 * We're now finished for good with this ioend structure.  Update the page
1041 * state, release holds on bios, and finally free up memory.  Do not use the
1042 * ioend after this.
1043 */
1044static void
1045iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1046{
1047	struct inode *inode = ioend->io_inode;
1048	struct bio *bio = &ioend->io_inline_bio;
1049	struct bio *last = ioend->io_bio, *next;
1050	u64 start = bio->bi_iter.bi_sector;
1051	loff_t offset = ioend->io_offset;
1052	bool quiet = bio_flagged(bio, BIO_QUIET);
1053
1054	for (bio = &ioend->io_inline_bio; bio; bio = next) {
1055		struct bio_vec *bv;
1056		struct bvec_iter_all iter_all;
1057
1058		/*
1059		 * For the last bio, bi_private points to the ioend, so we
1060		 * need to explicitly end the iteration here.
1061		 */
1062		if (bio == last)
1063			next = NULL;
1064		else
1065			next = bio->bi_private;
1066
1067		/* walk each page on bio, ending page IO on them */
1068		bio_for_each_segment_all(bv, bio, iter_all)
1069			iomap_finish_page_writeback(inode, bv->bv_page, error,
1070					bv->bv_len);
1071		bio_put(bio);
1072	}
1073	/* The ioend has been freed by bio_put() */
1074
1075	if (unlikely(error && !quiet)) {
1076		printk_ratelimited(KERN_ERR
1077"%s: writeback error on inode %lu, offset %lld, sector %llu",
1078			inode->i_sb->s_id, inode->i_ino, offset, start);
1079	}
1080}
1081
1082void
1083iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1084{
1085	struct list_head tmp;
1086
1087	list_replace_init(&ioend->io_list, &tmp);
1088	iomap_finish_ioend(ioend, error);
1089
1090	while (!list_empty(&tmp)) {
1091		ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1092		list_del_init(&ioend->io_list);
1093		iomap_finish_ioend(ioend, error);
1094	}
1095}
1096EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1097
1098/*
1099 * We can merge two adjacent ioends if they have the same set of work to do.
1100 */
1101static bool
1102iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1103{
1104	if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1105		return false;
1106	if ((ioend->io_flags & IOMAP_F_SHARED) ^
1107	    (next->io_flags & IOMAP_F_SHARED))
1108		return false;
1109	if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1110	    (next->io_type == IOMAP_UNWRITTEN))
1111		return false;
1112	if (ioend->io_offset + ioend->io_size != next->io_offset)
1113		return false;
1114	return true;
1115}
1116
1117void
1118iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1119{
1120	struct iomap_ioend *next;
1121
1122	INIT_LIST_HEAD(&ioend->io_list);
1123
1124	while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1125			io_list))) {
1126		if (!iomap_ioend_can_merge(ioend, next))
1127			break;
1128		list_move_tail(&next->io_list, &ioend->io_list);
1129		ioend->io_size += next->io_size;
1130	}
1131}
1132EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1133
1134static int
1135iomap_ioend_compare(void *priv, const struct list_head *a,
1136		const struct list_head *b)
1137{
1138	struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1139	struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1140
1141	if (ia->io_offset < ib->io_offset)
1142		return -1;
1143	if (ia->io_offset > ib->io_offset)
1144		return 1;
1145	return 0;
1146}
1147
1148void
1149iomap_sort_ioends(struct list_head *ioend_list)
1150{
1151	list_sort(NULL, ioend_list, iomap_ioend_compare);
1152}
1153EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1154
1155static void iomap_writepage_end_bio(struct bio *bio)
1156{
1157	struct iomap_ioend *ioend = bio->bi_private;
1158
1159	iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1160}
1161
1162/*
1163 * Submit the final bio for an ioend.
1164 *
1165 * If @error is non-zero, it means that we have a situation where some part of
1166 * the submission process has failed after we've marked pages for writeback
1167 * and unlocked them.  In this situation, we need to fail the bio instead of
1168 * submitting it.  This typically only happens on a filesystem shutdown.
1169 */
1170static int
1171iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1172		int error)
1173{
1174	ioend->io_bio->bi_private = ioend;
1175	ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1176
1177	if (wpc->ops->prepare_ioend)
1178		error = wpc->ops->prepare_ioend(ioend, error);
1179	if (error) {
1180		/*
1181		 * If we're failing the IO now, just mark the ioend with an
1182		 * error and finish it.  This will run IO completion immediately
1183		 * as there is only one reference to the ioend at this point in
1184		 * time.
1185		 */
1186		ioend->io_bio->bi_status = errno_to_blk_status(error);
1187		bio_endio(ioend->io_bio);
1188		return error;
1189	}
1190
1191	submit_bio(ioend->io_bio);
1192	return 0;
1193}
1194
1195static struct iomap_ioend *
1196iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1197		loff_t offset, sector_t sector, struct writeback_control *wbc)
1198{
1199	struct iomap_ioend *ioend;
1200	struct bio *bio;
1201
1202	bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS, &iomap_ioend_bioset);
1203	bio_set_dev(bio, wpc->iomap.bdev);
1204	bio->bi_iter.bi_sector = sector;
1205	bio->bi_opf = REQ_OP_WRITE | wbc_to_write_flags(wbc);
1206	bio->bi_write_hint = inode->i_write_hint;
1207	wbc_init_bio(wbc, bio);
1208
1209	ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1210	INIT_LIST_HEAD(&ioend->io_list);
1211	ioend->io_type = wpc->iomap.type;
1212	ioend->io_flags = wpc->iomap.flags;
1213	ioend->io_inode = inode;
1214	ioend->io_size = 0;
1215	ioend->io_offset = offset;
1216	ioend->io_bio = bio;
1217	return ioend;
1218}
1219
1220/*
1221 * Allocate a new bio, and chain the old bio to the new one.
1222 *
1223 * Note that we have to perform the chaining in this unintuitive order
1224 * so that the bi_private linkage is set up in the right direction for the
1225 * traversal in iomap_finish_ioend().
1226 */
1227static struct bio *
1228iomap_chain_bio(struct bio *prev)
1229{
1230	struct bio *new;
1231
1232	new = bio_alloc(GFP_NOFS, BIO_MAX_VECS);
1233	bio_copy_dev(new, prev);/* also copies over blkcg information */
1234	new->bi_iter.bi_sector = bio_end_sector(prev);
1235	new->bi_opf = prev->bi_opf;
1236	new->bi_write_hint = prev->bi_write_hint;
1237
1238	bio_chain(prev, new);
1239	bio_get(prev);		/* for iomap_finish_ioend */
1240	submit_bio(prev);
1241	return new;
1242}
1243
1244static bool
1245iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1246		sector_t sector)
1247{
1248	if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1249	    (wpc->ioend->io_flags & IOMAP_F_SHARED))
1250		return false;
1251	if (wpc->iomap.type != wpc->ioend->io_type)
1252		return false;
1253	if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1254		return false;
1255	if (sector != bio_end_sector(wpc->ioend->io_bio))
1256		return false;
1257	return true;
1258}
1259
1260/*
1261 * Test to see if we have an existing ioend structure that we could append to
1262 * first; otherwise finish off the current ioend and start another.
1263 */
1264static void
1265iomap_add_to_ioend(struct inode *inode, loff_t offset, struct page *page,
1266		struct iomap_page *iop, struct iomap_writepage_ctx *wpc,
1267		struct writeback_control *wbc, struct list_head *iolist)
1268{
1269	sector_t sector = iomap_sector(&wpc->iomap, offset);
1270	unsigned len = i_blocksize(inode);
1271	unsigned poff = offset & (PAGE_SIZE - 1);
1272
1273	if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, offset, sector)) {
1274		if (wpc->ioend)
1275			list_add(&wpc->ioend->io_list, iolist);
1276		wpc->ioend = iomap_alloc_ioend(inode, wpc, offset, sector, wbc);
1277	}
1278
1279	if (bio_add_page(wpc->ioend->io_bio, page, len, poff) != len) {
1280		wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1281		__bio_add_page(wpc->ioend->io_bio, page, len, poff);
1282	}
1283
1284	if (iop)
1285		atomic_add(len, &iop->write_bytes_pending);
1286	wpc->ioend->io_size += len;
1287	wbc_account_cgroup_owner(wbc, page, len);
1288}
1289
1290/*
1291 * We implement an immediate ioend submission policy here to avoid needing to
1292 * chain multiple ioends and hence nest mempool allocations which can violate
1293 * the forward progress guarantees we need to provide. The current ioend we're
1294 * adding blocks to is cached in the writepage context, and if the new block
1295 * doesn't append to the cached ioend, it will create a new ioend and cache that
1296 * instead.
1297 *
1298 * If a new ioend is created and cached, the old ioend is returned and queued
1299 * locally for submission once the entire page is processed or an error has been
1300 * detected.  While ioends are submitted immediately after they are completed,
1301 * batching optimisations are provided by higher level block plugging.
1302 *
1303 * At the end of a writeback pass, there will be a cached ioend remaining on the
1304 * writepage context that the caller will need to submit.
1305 */
1306static int
1307iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1308		struct writeback_control *wbc, struct inode *inode,
1309		struct page *page, u64 end_offset)
1310{
1311	struct iomap_page *iop = iomap_page_create(inode, page);
1312	struct iomap_ioend *ioend, *next;
1313	unsigned len = i_blocksize(inode);
1314	u64 file_offset; /* file offset of page */
1315	int error = 0, count = 0, i;
1316	LIST_HEAD(submit_list);
1317
1318	WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) != 0);
1319
1320	/*
1321	 * Walk through the page to find areas to write back. If we run off the
1322	 * end of the current map or find the current map invalid, grab a new
1323	 * one.
1324	 */
1325	for (i = 0, file_offset = page_offset(page);
1326	     i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
1327	     i++, file_offset += len) {
1328		if (iop && !test_bit(i, iop->uptodate))
1329			continue;
1330
1331		error = wpc->ops->map_blocks(wpc, inode, file_offset);
1332		if (error)
1333			break;
1334		if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1335			continue;
1336		if (wpc->iomap.type == IOMAP_HOLE)
1337			continue;
1338		iomap_add_to_ioend(inode, file_offset, page, iop, wpc, wbc,
1339				 &submit_list);
1340		count++;
1341	}
1342
1343	WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1344	WARN_ON_ONCE(!PageLocked(page));
1345	WARN_ON_ONCE(PageWriteback(page));
1346	WARN_ON_ONCE(PageDirty(page));
1347
1348	/*
1349	 * We cannot cancel the ioend directly here on error.  We may have
1350	 * already set other pages under writeback and hence we have to run I/O
1351	 * completion to mark the error state of the pages under writeback
1352	 * appropriately.
1353	 */
1354	if (unlikely(error)) {
1355		/*
1356		 * Let the filesystem know what portion of the current page
1357		 * failed to map. If the page hasn't been added to ioend, it
1358		 * won't be affected by I/O completion and we must unlock it
1359		 * now.
1360		 */
1361		if (wpc->ops->discard_page)
1362			wpc->ops->discard_page(page, file_offset);
1363		if (!count) {
1364			ClearPageUptodate(page);
1365			unlock_page(page);
1366			goto done;
1367		}
1368	}
1369
1370	set_page_writeback(page);
1371	unlock_page(page);
1372
1373	/*
1374	 * Preserve the original error if there was one; catch
1375	 * submission errors here and propagate into subsequent ioend
1376	 * submissions.
1377	 */
1378	list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1379		int error2;
1380
1381		list_del_init(&ioend->io_list);
1382		error2 = iomap_submit_ioend(wpc, ioend, error);
1383		if (error2 && !error)
1384			error = error2;
1385	}
1386
1387	/*
1388	 * We can end up here with no error and nothing to write only if we race
1389	 * with a partial page truncate on a sub-page block sized filesystem.
1390	 */
1391	if (!count)
1392		end_page_writeback(page);
1393done:
1394	mapping_set_error(page->mapping, error);
1395	return error;
1396}
1397
1398/*
1399 * Write out a dirty page.
1400 *
1401 * For delalloc space on the page, we need to allocate space and flush it.
1402 * For unwritten space on the page, we need to start the conversion to
1403 * regular allocated space.
1404 */
1405static int
1406iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
1407{
1408	struct iomap_writepage_ctx *wpc = data;
1409	struct inode *inode = page->mapping->host;
1410	pgoff_t end_index;
1411	u64 end_offset;
1412	loff_t offset;
1413
1414	trace_iomap_writepage(inode, page_offset(page), PAGE_SIZE);
1415
1416	/*
1417	 * Refuse to write the page out if we're called from reclaim context.
1418	 *
1419	 * This avoids stack overflows when called from deeply used stacks in
1420	 * random callers for direct reclaim or memcg reclaim.  We explicitly
1421	 * allow reclaim from kswapd as the stack usage there is relatively low.
1422	 *
1423	 * This should never happen except in the case of a VM regression so
1424	 * warn about it.
1425	 */
1426	if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1427			PF_MEMALLOC))
1428		goto redirty;
1429
1430	/*
1431	 * Is this page beyond the end of the file?
1432	 *
1433	 * The page index is less than the end_index, adjust the end_offset
1434	 * to the highest offset that this page should represent.
1435	 * -----------------------------------------------------
1436	 * |			file mapping	       | <EOF> |
1437	 * -----------------------------------------------------
1438	 * | Page ... | Page N-2 | Page N-1 |  Page N  |       |
1439	 * ^--------------------------------^----------|--------
1440	 * |     desired writeback range    |      see else    |
1441	 * ---------------------------------^------------------|
1442	 */
1443	offset = i_size_read(inode);
1444	end_index = offset >> PAGE_SHIFT;
1445	if (page->index < end_index)
1446		end_offset = (loff_t)(page->index + 1) << PAGE_SHIFT;
1447	else {
1448		/*
1449		 * Check whether the page to write out is beyond or straddles
1450		 * i_size or not.
1451		 * -------------------------------------------------------
1452		 * |		file mapping		        | <EOF>  |
1453		 * -------------------------------------------------------
1454		 * | Page ... | Page N-2 | Page N-1 |  Page N   | Beyond |
1455		 * ^--------------------------------^-----------|---------
1456		 * |				    |      Straddles     |
1457		 * ---------------------------------^-----------|--------|
1458		 */
1459		unsigned offset_into_page = offset & (PAGE_SIZE - 1);
1460
1461		/*
1462		 * Skip the page if it's fully outside i_size, e.g. due to a
1463		 * truncate operation that's in progress. We must redirty the
1464		 * page so that reclaim stops reclaiming it. Otherwise
1465		 * iomap_vm_releasepage() is called on it and gets confused.
1466		 *
1467		 * Note that the end_index is unsigned long.  If the given
1468		 * offset is greater than 16TB on a 32-bit system then if we
1469		 * checked if the page is fully outside i_size with
1470		 * "if (page->index >= end_index + 1)", "end_index + 1" would
1471		 * overflow and evaluate to 0.  Hence this page would be
1472		 * redirtied and written out repeatedly, which would result in
1473		 * an infinite loop; the user program performing this operation
1474		 * would hang.  Instead, we can detect this situation by
1475		 * checking if the page is totally beyond i_size or if its
1476		 * offset is just equal to the EOF.
1477		 */
1478		if (page->index > end_index ||
1479		    (page->index == end_index && offset_into_page == 0))
1480			goto redirty;
1481
1482		/*
1483		 * The page straddles i_size.  It must be zeroed out on each
1484		 * and every writepage invocation because it may be mmapped.
1485		 * "A file is mapped in multiples of the page size.  For a file
1486		 * that is not a multiple of the page size, the remaining
1487		 * memory is zeroed when mapped, and writes to that region are
1488		 * not written out to the file."
1489		 */
1490		zero_user_segment(page, offset_into_page, PAGE_SIZE);
1491
1492		/* Adjust the end_offset to the end of file */
1493		end_offset = offset;
1494	}
1495
1496	return iomap_writepage_map(wpc, wbc, inode, page, end_offset);
1497
1498redirty:
1499	redirty_page_for_writepage(wbc, page);
1500	unlock_page(page);
1501	return 0;
1502}
1503
1504int
1505iomap_writepage(struct page *page, struct writeback_control *wbc,
1506		struct iomap_writepage_ctx *wpc,
1507		const struct iomap_writeback_ops *ops)
1508{
1509	int ret;
1510
1511	wpc->ops = ops;
1512	ret = iomap_do_writepage(page, wbc, wpc);
1513	if (!wpc->ioend)
1514		return ret;
1515	return iomap_submit_ioend(wpc, wpc->ioend, ret);
1516}
1517EXPORT_SYMBOL_GPL(iomap_writepage);
1518
1519int
1520iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1521		struct iomap_writepage_ctx *wpc,
1522		const struct iomap_writeback_ops *ops)
1523{
1524	int			ret;
1525
1526	wpc->ops = ops;
1527	ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1528	if (!wpc->ioend)
1529		return ret;
1530	return iomap_submit_ioend(wpc, wpc->ioend, ret);
1531}
1532EXPORT_SYMBOL_GPL(iomap_writepages);
1533
1534static int __init iomap_init(void)
1535{
1536	return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
1537			   offsetof(struct iomap_ioend, io_inline_bio),
1538			   BIOSET_NEED_BVECS);
1539}
1540fs_initcall(iomap_init);
1541