xref: /linux/fs/iomap/buffered-io.c (revision a0efa2f362a69e47b9d8b48f770ef3a0249a7911)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2010 Red Hat, Inc.
4  * Copyright (C) 2016-2023 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 #define IOEND_BATCH_SIZE	4096
25 
26 /*
27  * Structure allocated for each folio to track per-block uptodate, dirty state
28  * and I/O completions.
29  */
30 struct iomap_folio_state {
31 	spinlock_t		state_lock;
32 	unsigned int		read_bytes_pending;
33 	atomic_t		write_bytes_pending;
34 
35 	/*
36 	 * Each block has two bits in this bitmap:
37 	 * Bits [0..blocks_per_folio) has the uptodate status.
38 	 * Bits [b_p_f...(2*b_p_f))   has the dirty status.
39 	 */
40 	unsigned long		state[];
41 };
42 
43 static struct bio_set iomap_ioend_bioset;
44 
45 static inline bool ifs_is_fully_uptodate(struct folio *folio,
46 		struct iomap_folio_state *ifs)
47 {
48 	struct inode *inode = folio->mapping->host;
49 
50 	return bitmap_full(ifs->state, i_blocks_per_folio(inode, folio));
51 }
52 
53 static inline bool ifs_block_is_uptodate(struct iomap_folio_state *ifs,
54 		unsigned int block)
55 {
56 	return test_bit(block, ifs->state);
57 }
58 
59 static bool ifs_set_range_uptodate(struct folio *folio,
60 		struct iomap_folio_state *ifs, size_t off, size_t len)
61 {
62 	struct inode *inode = folio->mapping->host;
63 	unsigned int first_blk = off >> inode->i_blkbits;
64 	unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
65 	unsigned int nr_blks = last_blk - first_blk + 1;
66 
67 	bitmap_set(ifs->state, first_blk, nr_blks);
68 	return ifs_is_fully_uptodate(folio, ifs);
69 }
70 
71 static void iomap_set_range_uptodate(struct folio *folio, size_t off,
72 		size_t len)
73 {
74 	struct iomap_folio_state *ifs = folio->private;
75 	unsigned long flags;
76 	bool uptodate = true;
77 
78 	if (ifs) {
79 		spin_lock_irqsave(&ifs->state_lock, flags);
80 		uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
81 		spin_unlock_irqrestore(&ifs->state_lock, flags);
82 	}
83 
84 	if (uptodate)
85 		folio_mark_uptodate(folio);
86 }
87 
88 static inline bool ifs_block_is_dirty(struct folio *folio,
89 		struct iomap_folio_state *ifs, int block)
90 {
91 	struct inode *inode = folio->mapping->host;
92 	unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
93 
94 	return test_bit(block + blks_per_folio, ifs->state);
95 }
96 
97 static unsigned ifs_find_dirty_range(struct folio *folio,
98 		struct iomap_folio_state *ifs, u64 *range_start, u64 range_end)
99 {
100 	struct inode *inode = folio->mapping->host;
101 	unsigned start_blk =
102 		offset_in_folio(folio, *range_start) >> inode->i_blkbits;
103 	unsigned end_blk = min_not_zero(
104 		offset_in_folio(folio, range_end) >> inode->i_blkbits,
105 		i_blocks_per_folio(inode, folio));
106 	unsigned nblks = 1;
107 
108 	while (!ifs_block_is_dirty(folio, ifs, start_blk))
109 		if (++start_blk == end_blk)
110 			return 0;
111 
112 	while (start_blk + nblks < end_blk) {
113 		if (!ifs_block_is_dirty(folio, ifs, start_blk + nblks))
114 			break;
115 		nblks++;
116 	}
117 
118 	*range_start = folio_pos(folio) + (start_blk << inode->i_blkbits);
119 	return nblks << inode->i_blkbits;
120 }
121 
122 static unsigned iomap_find_dirty_range(struct folio *folio, u64 *range_start,
123 		u64 range_end)
124 {
125 	struct iomap_folio_state *ifs = folio->private;
126 
127 	if (*range_start >= range_end)
128 		return 0;
129 
130 	if (ifs)
131 		return ifs_find_dirty_range(folio, ifs, range_start, range_end);
132 	return range_end - *range_start;
133 }
134 
135 static void ifs_clear_range_dirty(struct folio *folio,
136 		struct iomap_folio_state *ifs, size_t off, size_t len)
137 {
138 	struct inode *inode = folio->mapping->host;
139 	unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
140 	unsigned int first_blk = (off >> inode->i_blkbits);
141 	unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
142 	unsigned int nr_blks = last_blk - first_blk + 1;
143 	unsigned long flags;
144 
145 	spin_lock_irqsave(&ifs->state_lock, flags);
146 	bitmap_clear(ifs->state, first_blk + blks_per_folio, nr_blks);
147 	spin_unlock_irqrestore(&ifs->state_lock, flags);
148 }
149 
150 static void iomap_clear_range_dirty(struct folio *folio, size_t off, size_t len)
151 {
152 	struct iomap_folio_state *ifs = folio->private;
153 
154 	if (ifs)
155 		ifs_clear_range_dirty(folio, ifs, off, len);
156 }
157 
158 static void ifs_set_range_dirty(struct folio *folio,
159 		struct iomap_folio_state *ifs, size_t off, size_t len)
160 {
161 	struct inode *inode = folio->mapping->host;
162 	unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
163 	unsigned int first_blk = (off >> inode->i_blkbits);
164 	unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
165 	unsigned int nr_blks = last_blk - first_blk + 1;
166 	unsigned long flags;
167 
168 	spin_lock_irqsave(&ifs->state_lock, flags);
169 	bitmap_set(ifs->state, first_blk + blks_per_folio, nr_blks);
170 	spin_unlock_irqrestore(&ifs->state_lock, flags);
171 }
172 
173 static void iomap_set_range_dirty(struct folio *folio, size_t off, size_t len)
174 {
175 	struct iomap_folio_state *ifs = folio->private;
176 
177 	if (ifs)
178 		ifs_set_range_dirty(folio, ifs, off, len);
179 }
180 
181 static struct iomap_folio_state *ifs_alloc(struct inode *inode,
182 		struct folio *folio, unsigned int flags)
183 {
184 	struct iomap_folio_state *ifs = folio->private;
185 	unsigned int nr_blocks = i_blocks_per_folio(inode, folio);
186 	gfp_t gfp;
187 
188 	if (ifs || nr_blocks <= 1)
189 		return ifs;
190 
191 	if (flags & IOMAP_NOWAIT)
192 		gfp = GFP_NOWAIT;
193 	else
194 		gfp = GFP_NOFS | __GFP_NOFAIL;
195 
196 	/*
197 	 * ifs->state tracks two sets of state flags when the
198 	 * filesystem block size is smaller than the folio size.
199 	 * The first state tracks per-block uptodate and the
200 	 * second tracks per-block dirty state.
201 	 */
202 	ifs = kzalloc(struct_size(ifs, state,
203 		      BITS_TO_LONGS(2 * nr_blocks)), gfp);
204 	if (!ifs)
205 		return ifs;
206 
207 	spin_lock_init(&ifs->state_lock);
208 	if (folio_test_uptodate(folio))
209 		bitmap_set(ifs->state, 0, nr_blocks);
210 	if (folio_test_dirty(folio))
211 		bitmap_set(ifs->state, nr_blocks, nr_blocks);
212 	folio_attach_private(folio, ifs);
213 
214 	return ifs;
215 }
216 
217 static void ifs_free(struct folio *folio)
218 {
219 	struct iomap_folio_state *ifs = folio_detach_private(folio);
220 
221 	if (!ifs)
222 		return;
223 	WARN_ON_ONCE(ifs->read_bytes_pending != 0);
224 	WARN_ON_ONCE(atomic_read(&ifs->write_bytes_pending));
225 	WARN_ON_ONCE(ifs_is_fully_uptodate(folio, ifs) !=
226 			folio_test_uptodate(folio));
227 	kfree(ifs);
228 }
229 
230 /*
231  * Calculate the range inside the folio that we actually need to read.
232  */
233 static void iomap_adjust_read_range(struct inode *inode, struct folio *folio,
234 		loff_t *pos, loff_t length, size_t *offp, size_t *lenp)
235 {
236 	struct iomap_folio_state *ifs = folio->private;
237 	loff_t orig_pos = *pos;
238 	loff_t isize = i_size_read(inode);
239 	unsigned block_bits = inode->i_blkbits;
240 	unsigned block_size = (1 << block_bits);
241 	size_t poff = offset_in_folio(folio, *pos);
242 	size_t plen = min_t(loff_t, folio_size(folio) - poff, length);
243 	size_t orig_plen = plen;
244 	unsigned first = poff >> block_bits;
245 	unsigned last = (poff + plen - 1) >> block_bits;
246 
247 	/*
248 	 * If the block size is smaller than the page size, we need to check the
249 	 * per-block uptodate status and adjust the offset and length if needed
250 	 * to avoid reading in already uptodate ranges.
251 	 */
252 	if (ifs) {
253 		unsigned int i;
254 
255 		/* move forward for each leading block marked uptodate */
256 		for (i = first; i <= last; i++) {
257 			if (!ifs_block_is_uptodate(ifs, i))
258 				break;
259 			*pos += block_size;
260 			poff += block_size;
261 			plen -= block_size;
262 			first++;
263 		}
264 
265 		/* truncate len if we find any trailing uptodate block(s) */
266 		for ( ; i <= last; i++) {
267 			if (ifs_block_is_uptodate(ifs, i)) {
268 				plen -= (last - i + 1) * block_size;
269 				last = i - 1;
270 				break;
271 			}
272 		}
273 	}
274 
275 	/*
276 	 * If the extent spans the block that contains the i_size, we need to
277 	 * handle both halves separately so that we properly zero data in the
278 	 * page cache for blocks that are entirely outside of i_size.
279 	 */
280 	if (orig_pos <= isize && orig_pos + orig_plen > isize) {
281 		unsigned end = offset_in_folio(folio, isize - 1) >> block_bits;
282 
283 		if (first <= end && last > end)
284 			plen -= (last - end) * block_size;
285 	}
286 
287 	*offp = poff;
288 	*lenp = plen;
289 }
290 
291 static void iomap_finish_folio_read(struct folio *folio, size_t off,
292 		size_t len, int error)
293 {
294 	struct iomap_folio_state *ifs = folio->private;
295 	bool uptodate = !error;
296 	bool finished = true;
297 
298 	if (ifs) {
299 		unsigned long flags;
300 
301 		spin_lock_irqsave(&ifs->state_lock, flags);
302 		if (!error)
303 			uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
304 		ifs->read_bytes_pending -= len;
305 		finished = !ifs->read_bytes_pending;
306 		spin_unlock_irqrestore(&ifs->state_lock, flags);
307 	}
308 
309 	if (finished)
310 		folio_end_read(folio, uptodate);
311 }
312 
313 static void iomap_read_end_io(struct bio *bio)
314 {
315 	int error = blk_status_to_errno(bio->bi_status);
316 	struct folio_iter fi;
317 
318 	bio_for_each_folio_all(fi, bio)
319 		iomap_finish_folio_read(fi.folio, fi.offset, fi.length, error);
320 	bio_put(bio);
321 }
322 
323 struct iomap_readpage_ctx {
324 	struct folio		*cur_folio;
325 	bool			cur_folio_in_bio;
326 	struct bio		*bio;
327 	struct readahead_control *rac;
328 };
329 
330 /**
331  * iomap_read_inline_data - copy inline data into the page cache
332  * @iter: iteration structure
333  * @folio: folio to copy to
334  *
335  * Copy the inline data in @iter into @folio and zero out the rest of the folio.
336  * Only a single IOMAP_INLINE extent is allowed at the end of each file.
337  * Returns zero for success to complete the read, or the usual negative errno.
338  */
339 static int iomap_read_inline_data(const struct iomap_iter *iter,
340 		struct folio *folio)
341 {
342 	const struct iomap *iomap = iomap_iter_srcmap(iter);
343 	size_t size = i_size_read(iter->inode) - iomap->offset;
344 	size_t offset = offset_in_folio(folio, iomap->offset);
345 
346 	if (folio_test_uptodate(folio))
347 		return 0;
348 
349 	if (WARN_ON_ONCE(size > iomap->length))
350 		return -EIO;
351 	if (offset > 0)
352 		ifs_alloc(iter->inode, folio, iter->flags);
353 
354 	folio_fill_tail(folio, offset, iomap->inline_data, size);
355 	iomap_set_range_uptodate(folio, offset, folio_size(folio) - offset);
356 	return 0;
357 }
358 
359 static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
360 		loff_t pos)
361 {
362 	const struct iomap *srcmap = iomap_iter_srcmap(iter);
363 
364 	return srcmap->type != IOMAP_MAPPED ||
365 		(srcmap->flags & IOMAP_F_NEW) ||
366 		pos >= i_size_read(iter->inode);
367 }
368 
369 static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
370 		struct iomap_readpage_ctx *ctx, loff_t offset)
371 {
372 	const struct iomap *iomap = &iter->iomap;
373 	loff_t pos = iter->pos + offset;
374 	loff_t length = iomap_length(iter) - offset;
375 	struct folio *folio = ctx->cur_folio;
376 	struct iomap_folio_state *ifs;
377 	loff_t orig_pos = pos;
378 	size_t poff, plen;
379 	sector_t sector;
380 
381 	if (iomap->type == IOMAP_INLINE)
382 		return iomap_read_inline_data(iter, folio);
383 
384 	/* zero post-eof blocks as the page may be mapped */
385 	ifs = ifs_alloc(iter->inode, folio, iter->flags);
386 	iomap_adjust_read_range(iter->inode, folio, &pos, length, &poff, &plen);
387 	if (plen == 0)
388 		goto done;
389 
390 	if (iomap_block_needs_zeroing(iter, pos)) {
391 		folio_zero_range(folio, poff, plen);
392 		iomap_set_range_uptodate(folio, poff, plen);
393 		goto done;
394 	}
395 
396 	ctx->cur_folio_in_bio = true;
397 	if (ifs) {
398 		spin_lock_irq(&ifs->state_lock);
399 		ifs->read_bytes_pending += plen;
400 		spin_unlock_irq(&ifs->state_lock);
401 	}
402 
403 	sector = iomap_sector(iomap, pos);
404 	if (!ctx->bio ||
405 	    bio_end_sector(ctx->bio) != sector ||
406 	    !bio_add_folio(ctx->bio, folio, plen, poff)) {
407 		gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
408 		gfp_t orig_gfp = gfp;
409 		unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
410 
411 		if (ctx->bio)
412 			submit_bio(ctx->bio);
413 
414 		if (ctx->rac) /* same as readahead_gfp_mask */
415 			gfp |= __GFP_NORETRY | __GFP_NOWARN;
416 		ctx->bio = bio_alloc(iomap->bdev, bio_max_segs(nr_vecs),
417 				     REQ_OP_READ, gfp);
418 		/*
419 		 * If the bio_alloc fails, try it again for a single page to
420 		 * avoid having to deal with partial page reads.  This emulates
421 		 * what do_mpage_read_folio does.
422 		 */
423 		if (!ctx->bio) {
424 			ctx->bio = bio_alloc(iomap->bdev, 1, REQ_OP_READ,
425 					     orig_gfp);
426 		}
427 		if (ctx->rac)
428 			ctx->bio->bi_opf |= REQ_RAHEAD;
429 		ctx->bio->bi_iter.bi_sector = sector;
430 		ctx->bio->bi_end_io = iomap_read_end_io;
431 		bio_add_folio_nofail(ctx->bio, folio, plen, poff);
432 	}
433 
434 done:
435 	/*
436 	 * Move the caller beyond our range so that it keeps making progress.
437 	 * For that, we have to include any leading non-uptodate ranges, but
438 	 * we can skip trailing ones as they will be handled in the next
439 	 * iteration.
440 	 */
441 	return pos - orig_pos + plen;
442 }
443 
444 static loff_t iomap_read_folio_iter(const struct iomap_iter *iter,
445 		struct iomap_readpage_ctx *ctx)
446 {
447 	struct folio *folio = ctx->cur_folio;
448 	size_t offset = offset_in_folio(folio, iter->pos);
449 	loff_t length = min_t(loff_t, folio_size(folio) - offset,
450 			      iomap_length(iter));
451 	loff_t done, ret;
452 
453 	for (done = 0; done < length; done += ret) {
454 		ret = iomap_readpage_iter(iter, ctx, done);
455 		if (ret <= 0)
456 			return ret;
457 	}
458 
459 	return done;
460 }
461 
462 int iomap_read_folio(struct folio *folio, const struct iomap_ops *ops)
463 {
464 	struct iomap_iter iter = {
465 		.inode		= folio->mapping->host,
466 		.pos		= folio_pos(folio),
467 		.len		= folio_size(folio),
468 	};
469 	struct iomap_readpage_ctx ctx = {
470 		.cur_folio	= folio,
471 	};
472 	int ret;
473 
474 	trace_iomap_readpage(iter.inode, 1);
475 
476 	while ((ret = iomap_iter(&iter, ops)) > 0)
477 		iter.processed = iomap_read_folio_iter(&iter, &ctx);
478 
479 	if (ctx.bio) {
480 		submit_bio(ctx.bio);
481 		WARN_ON_ONCE(!ctx.cur_folio_in_bio);
482 	} else {
483 		WARN_ON_ONCE(ctx.cur_folio_in_bio);
484 		folio_unlock(folio);
485 	}
486 
487 	/*
488 	 * Just like mpage_readahead and block_read_full_folio, we always
489 	 * return 0 and just set the folio error flag on errors.  This
490 	 * should be cleaned up throughout the stack eventually.
491 	 */
492 	return 0;
493 }
494 EXPORT_SYMBOL_GPL(iomap_read_folio);
495 
496 static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
497 		struct iomap_readpage_ctx *ctx)
498 {
499 	loff_t length = iomap_length(iter);
500 	loff_t done, ret;
501 
502 	for (done = 0; done < length; done += ret) {
503 		if (ctx->cur_folio &&
504 		    offset_in_folio(ctx->cur_folio, iter->pos + done) == 0) {
505 			if (!ctx->cur_folio_in_bio)
506 				folio_unlock(ctx->cur_folio);
507 			ctx->cur_folio = NULL;
508 		}
509 		if (!ctx->cur_folio) {
510 			ctx->cur_folio = readahead_folio(ctx->rac);
511 			ctx->cur_folio_in_bio = false;
512 		}
513 		ret = iomap_readpage_iter(iter, ctx, done);
514 		if (ret <= 0)
515 			return ret;
516 	}
517 
518 	return done;
519 }
520 
521 /**
522  * iomap_readahead - Attempt to read pages from a file.
523  * @rac: Describes the pages to be read.
524  * @ops: The operations vector for the filesystem.
525  *
526  * This function is for filesystems to call to implement their readahead
527  * address_space operation.
528  *
529  * Context: The @ops callbacks may submit I/O (eg to read the addresses of
530  * blocks from disc), and may wait for it.  The caller may be trying to
531  * access a different page, and so sleeping excessively should be avoided.
532  * It may allocate memory, but should avoid costly allocations.  This
533  * function is called with memalloc_nofs set, so allocations will not cause
534  * the filesystem to be reentered.
535  */
536 void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
537 {
538 	struct iomap_iter iter = {
539 		.inode	= rac->mapping->host,
540 		.pos	= readahead_pos(rac),
541 		.len	= readahead_length(rac),
542 	};
543 	struct iomap_readpage_ctx ctx = {
544 		.rac	= rac,
545 	};
546 
547 	trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
548 
549 	while (iomap_iter(&iter, ops) > 0)
550 		iter.processed = iomap_readahead_iter(&iter, &ctx);
551 
552 	if (ctx.bio)
553 		submit_bio(ctx.bio);
554 	if (ctx.cur_folio) {
555 		if (!ctx.cur_folio_in_bio)
556 			folio_unlock(ctx.cur_folio);
557 	}
558 }
559 EXPORT_SYMBOL_GPL(iomap_readahead);
560 
561 /*
562  * iomap_is_partially_uptodate checks whether blocks within a folio are
563  * uptodate or not.
564  *
565  * Returns true if all blocks which correspond to the specified part
566  * of the folio are uptodate.
567  */
568 bool iomap_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
569 {
570 	struct iomap_folio_state *ifs = folio->private;
571 	struct inode *inode = folio->mapping->host;
572 	unsigned first, last, i;
573 
574 	if (!ifs)
575 		return false;
576 
577 	/* Caller's range may extend past the end of this folio */
578 	count = min(folio_size(folio) - from, count);
579 
580 	/* First and last blocks in range within folio */
581 	first = from >> inode->i_blkbits;
582 	last = (from + count - 1) >> inode->i_blkbits;
583 
584 	for (i = first; i <= last; i++)
585 		if (!ifs_block_is_uptodate(ifs, i))
586 			return false;
587 	return true;
588 }
589 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
590 
591 /**
592  * iomap_get_folio - get a folio reference for writing
593  * @iter: iteration structure
594  * @pos: start offset of write
595  * @len: Suggested size of folio to create.
596  *
597  * Returns a locked reference to the folio at @pos, or an error pointer if the
598  * folio could not be obtained.
599  */
600 struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
601 {
602 	fgf_t fgp = FGP_WRITEBEGIN | FGP_NOFS;
603 
604 	if (iter->flags & IOMAP_NOWAIT)
605 		fgp |= FGP_NOWAIT;
606 	fgp |= fgf_set_order(len);
607 
608 	return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
609 			fgp, mapping_gfp_mask(iter->inode->i_mapping));
610 }
611 EXPORT_SYMBOL_GPL(iomap_get_folio);
612 
613 bool iomap_release_folio(struct folio *folio, gfp_t gfp_flags)
614 {
615 	trace_iomap_release_folio(folio->mapping->host, folio_pos(folio),
616 			folio_size(folio));
617 
618 	/*
619 	 * If the folio is dirty, we refuse to release our metadata because
620 	 * it may be partially dirty.  Once we track per-block dirty state,
621 	 * we can release the metadata if every block is dirty.
622 	 */
623 	if (folio_test_dirty(folio))
624 		return false;
625 	ifs_free(folio);
626 	return true;
627 }
628 EXPORT_SYMBOL_GPL(iomap_release_folio);
629 
630 void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
631 {
632 	trace_iomap_invalidate_folio(folio->mapping->host,
633 					folio_pos(folio) + offset, len);
634 
635 	/*
636 	 * If we're invalidating the entire folio, clear the dirty state
637 	 * from it and release it to avoid unnecessary buildup of the LRU.
638 	 */
639 	if (offset == 0 && len == folio_size(folio)) {
640 		WARN_ON_ONCE(folio_test_writeback(folio));
641 		folio_cancel_dirty(folio);
642 		ifs_free(folio);
643 	}
644 }
645 EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
646 
647 bool iomap_dirty_folio(struct address_space *mapping, struct folio *folio)
648 {
649 	struct inode *inode = mapping->host;
650 	size_t len = folio_size(folio);
651 
652 	ifs_alloc(inode, folio, 0);
653 	iomap_set_range_dirty(folio, 0, len);
654 	return filemap_dirty_folio(mapping, folio);
655 }
656 EXPORT_SYMBOL_GPL(iomap_dirty_folio);
657 
658 static void
659 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
660 {
661 	loff_t i_size = i_size_read(inode);
662 
663 	/*
664 	 * Only truncate newly allocated pages beyoned EOF, even if the
665 	 * write started inside the existing inode size.
666 	 */
667 	if (pos + len > i_size)
668 		truncate_pagecache_range(inode, max(pos, i_size),
669 					 pos + len - 1);
670 }
671 
672 static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
673 		size_t poff, size_t plen, const struct iomap *iomap)
674 {
675 	struct bio_vec bvec;
676 	struct bio bio;
677 
678 	bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
679 	bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
680 	bio_add_folio_nofail(&bio, folio, plen, poff);
681 	return submit_bio_wait(&bio);
682 }
683 
684 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
685 		size_t len, struct folio *folio)
686 {
687 	const struct iomap *srcmap = iomap_iter_srcmap(iter);
688 	struct iomap_folio_state *ifs;
689 	loff_t block_size = i_blocksize(iter->inode);
690 	loff_t block_start = round_down(pos, block_size);
691 	loff_t block_end = round_up(pos + len, block_size);
692 	unsigned int nr_blocks = i_blocks_per_folio(iter->inode, folio);
693 	size_t from = offset_in_folio(folio, pos), to = from + len;
694 	size_t poff, plen;
695 
696 	/*
697 	 * If the write or zeroing completely overlaps the current folio, then
698 	 * entire folio will be dirtied so there is no need for
699 	 * per-block state tracking structures to be attached to this folio.
700 	 * For the unshare case, we must read in the ondisk contents because we
701 	 * are not changing pagecache contents.
702 	 */
703 	if (!(iter->flags & IOMAP_UNSHARE) && pos <= folio_pos(folio) &&
704 	    pos + len >= folio_pos(folio) + folio_size(folio))
705 		return 0;
706 
707 	ifs = ifs_alloc(iter->inode, folio, iter->flags);
708 	if ((iter->flags & IOMAP_NOWAIT) && !ifs && nr_blocks > 1)
709 		return -EAGAIN;
710 
711 	if (folio_test_uptodate(folio))
712 		return 0;
713 
714 	do {
715 		iomap_adjust_read_range(iter->inode, folio, &block_start,
716 				block_end - block_start, &poff, &plen);
717 		if (plen == 0)
718 			break;
719 
720 		if (!(iter->flags & IOMAP_UNSHARE) &&
721 		    (from <= poff || from >= poff + plen) &&
722 		    (to <= poff || to >= poff + plen))
723 			continue;
724 
725 		if (iomap_block_needs_zeroing(iter, block_start)) {
726 			if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
727 				return -EIO;
728 			folio_zero_segments(folio, poff, from, to, poff + plen);
729 		} else {
730 			int status;
731 
732 			if (iter->flags & IOMAP_NOWAIT)
733 				return -EAGAIN;
734 
735 			status = iomap_read_folio_sync(block_start, folio,
736 					poff, plen, srcmap);
737 			if (status)
738 				return status;
739 		}
740 		iomap_set_range_uptodate(folio, poff, plen);
741 	} while ((block_start += plen) < block_end);
742 
743 	return 0;
744 }
745 
746 static struct folio *__iomap_get_folio(struct iomap_iter *iter, loff_t pos,
747 		size_t len)
748 {
749 	const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
750 
751 	if (folio_ops && folio_ops->get_folio)
752 		return folio_ops->get_folio(iter, pos, len);
753 	else
754 		return iomap_get_folio(iter, pos, len);
755 }
756 
757 static void __iomap_put_folio(struct iomap_iter *iter, loff_t pos, size_t ret,
758 		struct folio *folio)
759 {
760 	const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
761 
762 	if (folio_ops && folio_ops->put_folio) {
763 		folio_ops->put_folio(iter->inode, pos, ret, folio);
764 	} else {
765 		folio_unlock(folio);
766 		folio_put(folio);
767 	}
768 }
769 
770 static int iomap_write_begin_inline(const struct iomap_iter *iter,
771 		struct folio *folio)
772 {
773 	/* needs more work for the tailpacking case; disable for now */
774 	if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
775 		return -EIO;
776 	return iomap_read_inline_data(iter, folio);
777 }
778 
779 static int iomap_write_begin(struct iomap_iter *iter, loff_t pos,
780 		size_t len, struct folio **foliop)
781 {
782 	const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
783 	const struct iomap *srcmap = iomap_iter_srcmap(iter);
784 	struct folio *folio;
785 	int status = 0;
786 
787 	BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
788 	if (srcmap != &iter->iomap)
789 		BUG_ON(pos + len > srcmap->offset + srcmap->length);
790 
791 	if (fatal_signal_pending(current))
792 		return -EINTR;
793 
794 	if (!mapping_large_folio_support(iter->inode->i_mapping))
795 		len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
796 
797 	folio = __iomap_get_folio(iter, pos, len);
798 	if (IS_ERR(folio))
799 		return PTR_ERR(folio);
800 
801 	/*
802 	 * Now we have a locked folio, before we do anything with it we need to
803 	 * check that the iomap we have cached is not stale. The inode extent
804 	 * mapping can change due to concurrent IO in flight (e.g.
805 	 * IOMAP_UNWRITTEN state can change and memory reclaim could have
806 	 * reclaimed a previously partially written page at this index after IO
807 	 * completion before this write reaches this file offset) and hence we
808 	 * could do the wrong thing here (zero a page range incorrectly or fail
809 	 * to zero) and corrupt data.
810 	 */
811 	if (folio_ops && folio_ops->iomap_valid) {
812 		bool iomap_valid = folio_ops->iomap_valid(iter->inode,
813 							 &iter->iomap);
814 		if (!iomap_valid) {
815 			iter->iomap.flags |= IOMAP_F_STALE;
816 			status = 0;
817 			goto out_unlock;
818 		}
819 	}
820 
821 	if (pos + len > folio_pos(folio) + folio_size(folio))
822 		len = folio_pos(folio) + folio_size(folio) - pos;
823 
824 	if (srcmap->type == IOMAP_INLINE)
825 		status = iomap_write_begin_inline(iter, folio);
826 	else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
827 		status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
828 	else
829 		status = __iomap_write_begin(iter, pos, len, folio);
830 
831 	if (unlikely(status))
832 		goto out_unlock;
833 
834 	*foliop = folio;
835 	return 0;
836 
837 out_unlock:
838 	__iomap_put_folio(iter, pos, 0, folio);
839 
840 	return status;
841 }
842 
843 static bool __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
844 		size_t copied, struct folio *folio)
845 {
846 	flush_dcache_folio(folio);
847 
848 	/*
849 	 * The blocks that were entirely written will now be uptodate, so we
850 	 * don't have to worry about a read_folio reading them and overwriting a
851 	 * partial write.  However, if we've encountered a short write and only
852 	 * partially written into a block, it will not be marked uptodate, so a
853 	 * read_folio might come in and destroy our partial write.
854 	 *
855 	 * Do the simplest thing and just treat any short write to a
856 	 * non-uptodate page as a zero-length write, and force the caller to
857 	 * redo the whole thing.
858 	 */
859 	if (unlikely(copied < len && !folio_test_uptodate(folio)))
860 		return false;
861 	iomap_set_range_uptodate(folio, offset_in_folio(folio, pos), len);
862 	iomap_set_range_dirty(folio, offset_in_folio(folio, pos), copied);
863 	filemap_dirty_folio(inode->i_mapping, folio);
864 	return true;
865 }
866 
867 static void iomap_write_end_inline(const struct iomap_iter *iter,
868 		struct folio *folio, loff_t pos, size_t copied)
869 {
870 	const struct iomap *iomap = &iter->iomap;
871 	void *addr;
872 
873 	WARN_ON_ONCE(!folio_test_uptodate(folio));
874 	BUG_ON(!iomap_inline_data_valid(iomap));
875 
876 	flush_dcache_folio(folio);
877 	addr = kmap_local_folio(folio, pos);
878 	memcpy(iomap_inline_data(iomap, pos), addr, copied);
879 	kunmap_local(addr);
880 
881 	mark_inode_dirty(iter->inode);
882 }
883 
884 /*
885  * Returns true if all copied bytes have been written to the pagecache,
886  * otherwise return false.
887  */
888 static bool iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
889 		size_t copied, struct folio *folio)
890 {
891 	const struct iomap *srcmap = iomap_iter_srcmap(iter);
892 
893 	if (srcmap->type == IOMAP_INLINE) {
894 		iomap_write_end_inline(iter, folio, pos, copied);
895 		return true;
896 	}
897 
898 	if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
899 		size_t bh_written;
900 
901 		bh_written = block_write_end(NULL, iter->inode->i_mapping, pos,
902 					len, copied, folio, NULL);
903 		WARN_ON_ONCE(bh_written != copied && bh_written != 0);
904 		return bh_written == copied;
905 	}
906 
907 	return __iomap_write_end(iter->inode, pos, len, copied, folio);
908 }
909 
910 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
911 {
912 	loff_t length = iomap_length(iter);
913 	loff_t pos = iter->pos;
914 	ssize_t total_written = 0;
915 	long status = 0;
916 	struct address_space *mapping = iter->inode->i_mapping;
917 	size_t chunk = mapping_max_folio_size(mapping);
918 	unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
919 
920 	do {
921 		struct folio *folio;
922 		loff_t old_size;
923 		size_t offset;		/* Offset into folio */
924 		size_t bytes;		/* Bytes to write to folio */
925 		size_t copied;		/* Bytes copied from user */
926 		size_t written;		/* Bytes have been written */
927 
928 		bytes = iov_iter_count(i);
929 retry:
930 		offset = pos & (chunk - 1);
931 		bytes = min(chunk - offset, bytes);
932 		status = balance_dirty_pages_ratelimited_flags(mapping,
933 							       bdp_flags);
934 		if (unlikely(status))
935 			break;
936 
937 		if (bytes > length)
938 			bytes = length;
939 
940 		/*
941 		 * Bring in the user page that we'll copy from _first_.
942 		 * Otherwise there's a nasty deadlock on copying from the
943 		 * same page as we're writing to, without it being marked
944 		 * up-to-date.
945 		 *
946 		 * For async buffered writes the assumption is that the user
947 		 * page has already been faulted in. This can be optimized by
948 		 * faulting the user page.
949 		 */
950 		if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
951 			status = -EFAULT;
952 			break;
953 		}
954 
955 		status = iomap_write_begin(iter, pos, bytes, &folio);
956 		if (unlikely(status)) {
957 			iomap_write_failed(iter->inode, pos, bytes);
958 			break;
959 		}
960 		if (iter->iomap.flags & IOMAP_F_STALE)
961 			break;
962 
963 		offset = offset_in_folio(folio, pos);
964 		if (bytes > folio_size(folio) - offset)
965 			bytes = folio_size(folio) - offset;
966 
967 		if (mapping_writably_mapped(mapping))
968 			flush_dcache_folio(folio);
969 
970 		copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
971 		written = iomap_write_end(iter, pos, bytes, copied, folio) ?
972 			  copied : 0;
973 
974 		/*
975 		 * Update the in-memory inode size after copying the data into
976 		 * the page cache.  It's up to the file system to write the
977 		 * updated size to disk, preferably after I/O completion so that
978 		 * no stale data is exposed.  Only once that's done can we
979 		 * unlock and release the folio.
980 		 */
981 		old_size = iter->inode->i_size;
982 		if (pos + written > old_size) {
983 			i_size_write(iter->inode, pos + written);
984 			iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
985 		}
986 		__iomap_put_folio(iter, pos, written, folio);
987 
988 		if (old_size < pos)
989 			pagecache_isize_extended(iter->inode, old_size, pos);
990 
991 		cond_resched();
992 		if (unlikely(written == 0)) {
993 			/*
994 			 * A short copy made iomap_write_end() reject the
995 			 * thing entirely.  Might be memory poisoning
996 			 * halfway through, might be a race with munmap,
997 			 * might be severe memory pressure.
998 			 */
999 			iomap_write_failed(iter->inode, pos, bytes);
1000 			iov_iter_revert(i, copied);
1001 
1002 			if (chunk > PAGE_SIZE)
1003 				chunk /= 2;
1004 			if (copied) {
1005 				bytes = copied;
1006 				goto retry;
1007 			}
1008 		} else {
1009 			pos += written;
1010 			total_written += written;
1011 			length -= written;
1012 		}
1013 	} while (iov_iter_count(i) && length);
1014 
1015 	if (status == -EAGAIN) {
1016 		iov_iter_revert(i, total_written);
1017 		return -EAGAIN;
1018 	}
1019 	return total_written ? total_written : status;
1020 }
1021 
1022 ssize_t
1023 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
1024 		const struct iomap_ops *ops, void *private)
1025 {
1026 	struct iomap_iter iter = {
1027 		.inode		= iocb->ki_filp->f_mapping->host,
1028 		.pos		= iocb->ki_pos,
1029 		.len		= iov_iter_count(i),
1030 		.flags		= IOMAP_WRITE,
1031 		.private	= private,
1032 	};
1033 	ssize_t ret;
1034 
1035 	if (iocb->ki_flags & IOCB_NOWAIT)
1036 		iter.flags |= IOMAP_NOWAIT;
1037 
1038 	while ((ret = iomap_iter(&iter, ops)) > 0)
1039 		iter.processed = iomap_write_iter(&iter, i);
1040 
1041 	if (unlikely(iter.pos == iocb->ki_pos))
1042 		return ret;
1043 	ret = iter.pos - iocb->ki_pos;
1044 	iocb->ki_pos = iter.pos;
1045 	return ret;
1046 }
1047 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
1048 
1049 static void iomap_write_delalloc_ifs_punch(struct inode *inode,
1050 		struct folio *folio, loff_t start_byte, loff_t end_byte,
1051 		struct iomap *iomap, iomap_punch_t punch)
1052 {
1053 	unsigned int first_blk, last_blk, i;
1054 	loff_t last_byte;
1055 	u8 blkbits = inode->i_blkbits;
1056 	struct iomap_folio_state *ifs;
1057 
1058 	/*
1059 	 * When we have per-block dirty tracking, there can be
1060 	 * blocks within a folio which are marked uptodate
1061 	 * but not dirty. In that case it is necessary to punch
1062 	 * out such blocks to avoid leaking any delalloc blocks.
1063 	 */
1064 	ifs = folio->private;
1065 	if (!ifs)
1066 		return;
1067 
1068 	last_byte = min_t(loff_t, end_byte - 1,
1069 			folio_pos(folio) + folio_size(folio) - 1);
1070 	first_blk = offset_in_folio(folio, start_byte) >> blkbits;
1071 	last_blk = offset_in_folio(folio, last_byte) >> blkbits;
1072 	for (i = first_blk; i <= last_blk; i++) {
1073 		if (!ifs_block_is_dirty(folio, ifs, i))
1074 			punch(inode, folio_pos(folio) + (i << blkbits),
1075 				    1 << blkbits, iomap);
1076 	}
1077 }
1078 
1079 static void iomap_write_delalloc_punch(struct inode *inode, struct folio *folio,
1080 		loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1081 		struct iomap *iomap, iomap_punch_t punch)
1082 {
1083 	if (!folio_test_dirty(folio))
1084 		return;
1085 
1086 	/* if dirty, punch up to offset */
1087 	if (start_byte > *punch_start_byte) {
1088 		punch(inode, *punch_start_byte, start_byte - *punch_start_byte,
1089 				iomap);
1090 	}
1091 
1092 	/* Punch non-dirty blocks within folio */
1093 	iomap_write_delalloc_ifs_punch(inode, folio, start_byte, end_byte,
1094 			iomap, punch);
1095 
1096 	/*
1097 	 * Make sure the next punch start is correctly bound to
1098 	 * the end of this data range, not the end of the folio.
1099 	 */
1100 	*punch_start_byte = min_t(loff_t, end_byte,
1101 				folio_pos(folio) + folio_size(folio));
1102 }
1103 
1104 /*
1105  * Scan the data range passed to us for dirty page cache folios. If we find a
1106  * dirty folio, punch out the preceding range and update the offset from which
1107  * the next punch will start from.
1108  *
1109  * We can punch out storage reservations under clean pages because they either
1110  * contain data that has been written back - in which case the delalloc punch
1111  * over that range is a no-op - or they have been read faults in which case they
1112  * contain zeroes and we can remove the delalloc backing range and any new
1113  * writes to those pages will do the normal hole filling operation...
1114  *
1115  * This makes the logic simple: we only need to keep the delalloc extents only
1116  * over the dirty ranges of the page cache.
1117  *
1118  * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1119  * simplify range iterations.
1120  */
1121 static void iomap_write_delalloc_scan(struct inode *inode,
1122 		loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1123 		struct iomap *iomap, iomap_punch_t punch)
1124 {
1125 	while (start_byte < end_byte) {
1126 		struct folio	*folio;
1127 
1128 		/* grab locked page */
1129 		folio = filemap_lock_folio(inode->i_mapping,
1130 				start_byte >> PAGE_SHIFT);
1131 		if (IS_ERR(folio)) {
1132 			start_byte = ALIGN_DOWN(start_byte, PAGE_SIZE) +
1133 					PAGE_SIZE;
1134 			continue;
1135 		}
1136 
1137 		iomap_write_delalloc_punch(inode, folio, punch_start_byte,
1138 				start_byte, end_byte, iomap, punch);
1139 
1140 		/* move offset to start of next folio in range */
1141 		start_byte = folio_next_index(folio) << PAGE_SHIFT;
1142 		folio_unlock(folio);
1143 		folio_put(folio);
1144 	}
1145 }
1146 
1147 /*
1148  * Punch out all the delalloc blocks in the range given except for those that
1149  * have dirty data still pending in the page cache - those are going to be
1150  * written and so must still retain the delalloc backing for writeback.
1151  *
1152  * As we are scanning the page cache for data, we don't need to reimplement the
1153  * wheel - mapping_seek_hole_data() does exactly what we need to identify the
1154  * start and end of data ranges correctly even for sub-folio block sizes. This
1155  * byte range based iteration is especially convenient because it means we
1156  * don't have to care about variable size folios, nor where the start or end of
1157  * the data range lies within a folio, if they lie within the same folio or even
1158  * if there are multiple discontiguous data ranges within the folio.
1159  *
1160  * It should be noted that mapping_seek_hole_data() is not aware of EOF, and so
1161  * can return data ranges that exist in the cache beyond EOF. e.g. a page fault
1162  * spanning EOF will initialise the post-EOF data to zeroes and mark it up to
1163  * date. A write page fault can then mark it dirty. If we then fail a write()
1164  * beyond EOF into that up to date cached range, we allocate a delalloc block
1165  * beyond EOF and then have to punch it out. Because the range is up to date,
1166  * mapping_seek_hole_data() will return it, and we will skip the punch because
1167  * the folio is dirty. THis is incorrect - we always need to punch out delalloc
1168  * beyond EOF in this case as writeback will never write back and covert that
1169  * delalloc block beyond EOF. Hence we limit the cached data scan range to EOF,
1170  * resulting in always punching out the range from the EOF to the end of the
1171  * range the iomap spans.
1172  *
1173  * Intervals are of the form [start_byte, end_byte) (i.e. open ended) because it
1174  * matches the intervals returned by mapping_seek_hole_data(). i.e. SEEK_DATA
1175  * returns the start of a data range (start_byte), and SEEK_HOLE(start_byte)
1176  * returns the end of the data range (data_end). Using closed intervals would
1177  * require sprinkling this code with magic "+ 1" and "- 1" arithmetic and expose
1178  * the code to subtle off-by-one bugs....
1179  */
1180 static void iomap_write_delalloc_release(struct inode *inode, loff_t start_byte,
1181 		loff_t end_byte, unsigned flags, struct iomap *iomap,
1182 		iomap_punch_t punch)
1183 {
1184 	loff_t punch_start_byte = start_byte;
1185 	loff_t scan_end_byte = min(i_size_read(inode), end_byte);
1186 
1187 	/*
1188 	 * Lock the mapping to avoid races with page faults re-instantiating
1189 	 * folios and dirtying them via ->page_mkwrite whilst we walk the
1190 	 * cache and perform delalloc extent removal. Failing to do this can
1191 	 * leave dirty pages with no space reservation in the cache.
1192 	 */
1193 	filemap_invalidate_lock(inode->i_mapping);
1194 	while (start_byte < scan_end_byte) {
1195 		loff_t		data_end;
1196 
1197 		start_byte = mapping_seek_hole_data(inode->i_mapping,
1198 				start_byte, scan_end_byte, SEEK_DATA);
1199 		/*
1200 		 * If there is no more data to scan, all that is left is to
1201 		 * punch out the remaining range.
1202 		 *
1203 		 * Note that mapping_seek_hole_data is only supposed to return
1204 		 * either an offset or -ENXIO, so WARN on any other error as
1205 		 * that would be an API change without updating the callers.
1206 		 */
1207 		if (start_byte == -ENXIO || start_byte == scan_end_byte)
1208 			break;
1209 		if (WARN_ON_ONCE(start_byte < 0))
1210 			goto out_unlock;
1211 		WARN_ON_ONCE(start_byte < punch_start_byte);
1212 		WARN_ON_ONCE(start_byte > scan_end_byte);
1213 
1214 		/*
1215 		 * We find the end of this contiguous cached data range by
1216 		 * seeking from start_byte to the beginning of the next hole.
1217 		 */
1218 		data_end = mapping_seek_hole_data(inode->i_mapping, start_byte,
1219 				scan_end_byte, SEEK_HOLE);
1220 		if (WARN_ON_ONCE(data_end < 0))
1221 			goto out_unlock;
1222 
1223 		/*
1224 		 * If we race with post-direct I/O invalidation of the page cache,
1225 		 * there might be no data left at start_byte.
1226 		 */
1227 		if (data_end == start_byte)
1228 			continue;
1229 
1230 		WARN_ON_ONCE(data_end < start_byte);
1231 		WARN_ON_ONCE(data_end > scan_end_byte);
1232 
1233 		iomap_write_delalloc_scan(inode, &punch_start_byte, start_byte,
1234 				data_end, iomap, punch);
1235 
1236 		/* The next data search starts at the end of this one. */
1237 		start_byte = data_end;
1238 	}
1239 
1240 	if (punch_start_byte < end_byte)
1241 		punch(inode, punch_start_byte, end_byte - punch_start_byte,
1242 				iomap);
1243 out_unlock:
1244 	filemap_invalidate_unlock(inode->i_mapping);
1245 }
1246 
1247 /*
1248  * When a short write occurs, the filesystem may need to remove reserved space
1249  * that was allocated in ->iomap_begin from it's ->iomap_end method. For
1250  * filesystems that use delayed allocation, we need to punch out delalloc
1251  * extents from the range that are not dirty in the page cache. As the write can
1252  * race with page faults, there can be dirty pages over the delalloc extent
1253  * outside the range of a short write but still within the delalloc extent
1254  * allocated for this iomap.
1255  *
1256  * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1257  * simplify range iterations.
1258  *
1259  * The punch() callback *must* only punch delalloc extents in the range passed
1260  * to it. It must skip over all other types of extents in the range and leave
1261  * them completely unchanged. It must do this punch atomically with respect to
1262  * other extent modifications.
1263  *
1264  * The punch() callback may be called with a folio locked to prevent writeback
1265  * extent allocation racing at the edge of the range we are currently punching.
1266  * The locked folio may or may not cover the range being punched, so it is not
1267  * safe for the punch() callback to lock folios itself.
1268  *
1269  * Lock order is:
1270  *
1271  * inode->i_rwsem (shared or exclusive)
1272  *   inode->i_mapping->invalidate_lock (exclusive)
1273  *     folio_lock()
1274  *       ->punch
1275  *         internal filesystem allocation lock
1276  */
1277 void iomap_file_buffered_write_punch_delalloc(struct inode *inode,
1278 		loff_t pos, loff_t length, ssize_t written, unsigned flags,
1279 		struct iomap *iomap, iomap_punch_t punch)
1280 {
1281 	loff_t			start_byte;
1282 	loff_t			end_byte;
1283 	unsigned int		blocksize = i_blocksize(inode);
1284 
1285 	if (iomap->type != IOMAP_DELALLOC)
1286 		return;
1287 
1288 	/* If we didn't reserve the blocks, we're not allowed to punch them. */
1289 	if (!(iomap->flags & IOMAP_F_NEW))
1290 		return;
1291 
1292 	/*
1293 	 * start_byte refers to the first unused block after a short write. If
1294 	 * nothing was written, round offset down to point at the first block in
1295 	 * the range.
1296 	 */
1297 	if (unlikely(!written))
1298 		start_byte = round_down(pos, blocksize);
1299 	else
1300 		start_byte = round_up(pos + written, blocksize);
1301 	end_byte = round_up(pos + length, blocksize);
1302 
1303 	/* Nothing to do if we've written the entire delalloc extent */
1304 	if (start_byte >= end_byte)
1305 		return;
1306 
1307 	iomap_write_delalloc_release(inode, start_byte, end_byte, flags, iomap,
1308 			punch);
1309 }
1310 EXPORT_SYMBOL_GPL(iomap_file_buffered_write_punch_delalloc);
1311 
1312 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
1313 {
1314 	struct iomap *iomap = &iter->iomap;
1315 	loff_t pos = iter->pos;
1316 	loff_t length = iomap_length(iter);
1317 	loff_t written = 0;
1318 
1319 	/* Don't bother with blocks that are not shared to start with. */
1320 	if (!(iomap->flags & IOMAP_F_SHARED))
1321 		return length;
1322 
1323 	/*
1324 	 * Don't bother with delalloc reservations, holes or unwritten extents.
1325 	 *
1326 	 * Note that we use srcmap directly instead of iomap_iter_srcmap as
1327 	 * unsharing requires providing a separate source map, and the presence
1328 	 * of one is a good indicator that unsharing is needed, unlike
1329 	 * IOMAP_F_SHARED which can be set for any data that goes into the COW
1330 	 * fork for XFS.
1331 	 */
1332 	if (iter->srcmap.type == IOMAP_HOLE ||
1333 	    iter->srcmap.type == IOMAP_DELALLOC ||
1334 	    iter->srcmap.type == IOMAP_UNWRITTEN)
1335 		return length;
1336 
1337 	do {
1338 		struct folio *folio;
1339 		int status;
1340 		size_t offset;
1341 		size_t bytes = min_t(u64, SIZE_MAX, length);
1342 		bool ret;
1343 
1344 		status = iomap_write_begin(iter, pos, bytes, &folio);
1345 		if (unlikely(status))
1346 			return status;
1347 		if (iomap->flags & IOMAP_F_STALE)
1348 			break;
1349 
1350 		offset = offset_in_folio(folio, pos);
1351 		if (bytes > folio_size(folio) - offset)
1352 			bytes = folio_size(folio) - offset;
1353 
1354 		ret = iomap_write_end(iter, pos, bytes, bytes, folio);
1355 		__iomap_put_folio(iter, pos, bytes, folio);
1356 		if (WARN_ON_ONCE(!ret))
1357 			return -EIO;
1358 
1359 		cond_resched();
1360 
1361 		pos += bytes;
1362 		written += bytes;
1363 		length -= bytes;
1364 
1365 		balance_dirty_pages_ratelimited(iter->inode->i_mapping);
1366 	} while (length > 0);
1367 
1368 	return written;
1369 }
1370 
1371 int
1372 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1373 		const struct iomap_ops *ops)
1374 {
1375 	struct iomap_iter iter = {
1376 		.inode		= inode,
1377 		.pos		= pos,
1378 		.flags		= IOMAP_WRITE | IOMAP_UNSHARE,
1379 	};
1380 	loff_t size = i_size_read(inode);
1381 	int ret;
1382 
1383 	if (pos < 0 || pos >= size)
1384 		return 0;
1385 
1386 	iter.len = min(len, size - pos);
1387 	while ((ret = iomap_iter(&iter, ops)) > 0)
1388 		iter.processed = iomap_unshare_iter(&iter);
1389 	return ret;
1390 }
1391 EXPORT_SYMBOL_GPL(iomap_file_unshare);
1392 
1393 /*
1394  * Flush the remaining range of the iter and mark the current mapping stale.
1395  * This is used when zero range sees an unwritten mapping that may have had
1396  * dirty pagecache over it.
1397  */
1398 static inline int iomap_zero_iter_flush_and_stale(struct iomap_iter *i)
1399 {
1400 	struct address_space *mapping = i->inode->i_mapping;
1401 	loff_t end = i->pos + i->len - 1;
1402 
1403 	i->iomap.flags |= IOMAP_F_STALE;
1404 	return filemap_write_and_wait_range(mapping, i->pos, end);
1405 }
1406 
1407 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero,
1408 		bool *range_dirty)
1409 {
1410 	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1411 	loff_t pos = iter->pos;
1412 	loff_t length = iomap_length(iter);
1413 	loff_t written = 0;
1414 
1415 	/*
1416 	 * We must zero subranges of unwritten mappings that might be dirty in
1417 	 * pagecache from previous writes. We only know whether the entire range
1418 	 * was clean or not, however, and dirty folios may have been written
1419 	 * back or reclaimed at any point after mapping lookup.
1420 	 *
1421 	 * The easiest way to deal with this is to flush pagecache to trigger
1422 	 * any pending unwritten conversions and then grab the updated extents
1423 	 * from the fs. The flush may change the current mapping, so mark it
1424 	 * stale for the iterator to remap it for the next pass to handle
1425 	 * properly.
1426 	 *
1427 	 * Note that holes are treated the same as unwritten because zero range
1428 	 * is (ab)used for partial folio zeroing in some cases. Hole backed
1429 	 * post-eof ranges can be dirtied via mapped write and the flush
1430 	 * triggers writeback time post-eof zeroing.
1431 	 */
1432 	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN) {
1433 		if (*range_dirty) {
1434 			*range_dirty = false;
1435 			return iomap_zero_iter_flush_and_stale(iter);
1436 		}
1437 		/* range is clean and already zeroed, nothing to do */
1438 		return length;
1439 	}
1440 
1441 	do {
1442 		struct folio *folio;
1443 		int status;
1444 		size_t offset;
1445 		size_t bytes = min_t(u64, SIZE_MAX, length);
1446 		bool ret;
1447 
1448 		status = iomap_write_begin(iter, pos, bytes, &folio);
1449 		if (status)
1450 			return status;
1451 		if (iter->iomap.flags & IOMAP_F_STALE)
1452 			break;
1453 
1454 		offset = offset_in_folio(folio, pos);
1455 		if (bytes > folio_size(folio) - offset)
1456 			bytes = folio_size(folio) - offset;
1457 
1458 		folio_zero_range(folio, offset, bytes);
1459 		folio_mark_accessed(folio);
1460 
1461 		ret = iomap_write_end(iter, pos, bytes, bytes, folio);
1462 		__iomap_put_folio(iter, pos, bytes, folio);
1463 		if (WARN_ON_ONCE(!ret))
1464 			return -EIO;
1465 
1466 		pos += bytes;
1467 		length -= bytes;
1468 		written += bytes;
1469 	} while (length > 0);
1470 
1471 	if (did_zero)
1472 		*did_zero = true;
1473 	return written;
1474 }
1475 
1476 int
1477 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1478 		const struct iomap_ops *ops)
1479 {
1480 	struct iomap_iter iter = {
1481 		.inode		= inode,
1482 		.pos		= pos,
1483 		.len		= len,
1484 		.flags		= IOMAP_ZERO,
1485 	};
1486 	int ret;
1487 	bool range_dirty;
1488 
1489 	/*
1490 	 * Zero range wants to skip pre-zeroed (i.e. unwritten) mappings, but
1491 	 * pagecache must be flushed to ensure stale data from previous
1492 	 * buffered writes is not exposed. A flush is only required for certain
1493 	 * types of mappings, but checking pagecache after mapping lookup is
1494 	 * racy with writeback and reclaim.
1495 	 *
1496 	 * Therefore, check the entire range first and pass along whether any
1497 	 * part of it is dirty. If so and an underlying mapping warrants it,
1498 	 * flush the cache at that point. This trades off the occasional false
1499 	 * positive (and spurious flush, if the dirty data and mapping don't
1500 	 * happen to overlap) for simplicity in handling a relatively uncommon
1501 	 * situation.
1502 	 */
1503 	range_dirty = filemap_range_needs_writeback(inode->i_mapping,
1504 					pos, pos + len - 1);
1505 
1506 	while ((ret = iomap_iter(&iter, ops)) > 0)
1507 		iter.processed = iomap_zero_iter(&iter, did_zero, &range_dirty);
1508 	return ret;
1509 }
1510 EXPORT_SYMBOL_GPL(iomap_zero_range);
1511 
1512 int
1513 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1514 		const struct iomap_ops *ops)
1515 {
1516 	unsigned int blocksize = i_blocksize(inode);
1517 	unsigned int off = pos & (blocksize - 1);
1518 
1519 	/* Block boundary? Nothing to do */
1520 	if (!off)
1521 		return 0;
1522 	return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1523 }
1524 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1525 
1526 static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
1527 		struct folio *folio)
1528 {
1529 	loff_t length = iomap_length(iter);
1530 	int ret;
1531 
1532 	if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
1533 		ret = __block_write_begin_int(folio, iter->pos, length, NULL,
1534 					      &iter->iomap);
1535 		if (ret)
1536 			return ret;
1537 		block_commit_write(&folio->page, 0, length);
1538 	} else {
1539 		WARN_ON_ONCE(!folio_test_uptodate(folio));
1540 		folio_mark_dirty(folio);
1541 	}
1542 
1543 	return length;
1544 }
1545 
1546 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1547 {
1548 	struct iomap_iter iter = {
1549 		.inode		= file_inode(vmf->vma->vm_file),
1550 		.flags		= IOMAP_WRITE | IOMAP_FAULT,
1551 	};
1552 	struct folio *folio = page_folio(vmf->page);
1553 	ssize_t ret;
1554 
1555 	folio_lock(folio);
1556 	ret = folio_mkwrite_check_truncate(folio, iter.inode);
1557 	if (ret < 0)
1558 		goto out_unlock;
1559 	iter.pos = folio_pos(folio);
1560 	iter.len = ret;
1561 	while ((ret = iomap_iter(&iter, ops)) > 0)
1562 		iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1563 
1564 	if (ret < 0)
1565 		goto out_unlock;
1566 	folio_wait_stable(folio);
1567 	return VM_FAULT_LOCKED;
1568 out_unlock:
1569 	folio_unlock(folio);
1570 	return vmf_fs_error(ret);
1571 }
1572 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1573 
1574 static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1575 		size_t len)
1576 {
1577 	struct iomap_folio_state *ifs = folio->private;
1578 
1579 	WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !ifs);
1580 	WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) <= 0);
1581 
1582 	if (!ifs || atomic_sub_and_test(len, &ifs->write_bytes_pending))
1583 		folio_end_writeback(folio);
1584 }
1585 
1586 /*
1587  * We're now finished for good with this ioend structure.  Update the page
1588  * state, release holds on bios, and finally free up memory.  Do not use the
1589  * ioend after this.
1590  */
1591 static u32
1592 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1593 {
1594 	struct inode *inode = ioend->io_inode;
1595 	struct bio *bio = &ioend->io_bio;
1596 	struct folio_iter fi;
1597 	u32 folio_count = 0;
1598 
1599 	if (error) {
1600 		mapping_set_error(inode->i_mapping, error);
1601 		if (!bio_flagged(bio, BIO_QUIET)) {
1602 			pr_err_ratelimited(
1603 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1604 				inode->i_sb->s_id, inode->i_ino,
1605 				ioend->io_offset, ioend->io_sector);
1606 		}
1607 	}
1608 
1609 	/* walk all folios in bio, ending page IO on them */
1610 	bio_for_each_folio_all(fi, bio) {
1611 		iomap_finish_folio_write(inode, fi.folio, fi.length);
1612 		folio_count++;
1613 	}
1614 
1615 	bio_put(bio);	/* frees the ioend */
1616 	return folio_count;
1617 }
1618 
1619 /*
1620  * Ioend completion routine for merged bios. This can only be called from task
1621  * contexts as merged ioends can be of unbound length. Hence we have to break up
1622  * the writeback completions into manageable chunks to avoid long scheduler
1623  * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1624  * good batch processing throughput without creating adverse scheduler latency
1625  * conditions.
1626  */
1627 void
1628 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1629 {
1630 	struct list_head tmp;
1631 	u32 completions;
1632 
1633 	might_sleep();
1634 
1635 	list_replace_init(&ioend->io_list, &tmp);
1636 	completions = iomap_finish_ioend(ioend, error);
1637 
1638 	while (!list_empty(&tmp)) {
1639 		if (completions > IOEND_BATCH_SIZE * 8) {
1640 			cond_resched();
1641 			completions = 0;
1642 		}
1643 		ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1644 		list_del_init(&ioend->io_list);
1645 		completions += iomap_finish_ioend(ioend, error);
1646 	}
1647 }
1648 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1649 
1650 /*
1651  * We can merge two adjacent ioends if they have the same set of work to do.
1652  */
1653 static bool
1654 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1655 {
1656 	if (ioend->io_bio.bi_status != next->io_bio.bi_status)
1657 		return false;
1658 	if ((ioend->io_flags & IOMAP_F_SHARED) ^
1659 	    (next->io_flags & IOMAP_F_SHARED))
1660 		return false;
1661 	if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1662 	    (next->io_type == IOMAP_UNWRITTEN))
1663 		return false;
1664 	if (ioend->io_offset + ioend->io_size != next->io_offset)
1665 		return false;
1666 	/*
1667 	 * Do not merge physically discontiguous ioends. The filesystem
1668 	 * completion functions will have to iterate the physical
1669 	 * discontiguities even if we merge the ioends at a logical level, so
1670 	 * we don't gain anything by merging physical discontiguities here.
1671 	 *
1672 	 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1673 	 * submission so does not point to the start sector of the bio at
1674 	 * completion.
1675 	 */
1676 	if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1677 		return false;
1678 	return true;
1679 }
1680 
1681 void
1682 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1683 {
1684 	struct iomap_ioend *next;
1685 
1686 	INIT_LIST_HEAD(&ioend->io_list);
1687 
1688 	while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1689 			io_list))) {
1690 		if (!iomap_ioend_can_merge(ioend, next))
1691 			break;
1692 		list_move_tail(&next->io_list, &ioend->io_list);
1693 		ioend->io_size += next->io_size;
1694 	}
1695 }
1696 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1697 
1698 static int
1699 iomap_ioend_compare(void *priv, const struct list_head *a,
1700 		const struct list_head *b)
1701 {
1702 	struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1703 	struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1704 
1705 	if (ia->io_offset < ib->io_offset)
1706 		return -1;
1707 	if (ia->io_offset > ib->io_offset)
1708 		return 1;
1709 	return 0;
1710 }
1711 
1712 void
1713 iomap_sort_ioends(struct list_head *ioend_list)
1714 {
1715 	list_sort(NULL, ioend_list, iomap_ioend_compare);
1716 }
1717 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1718 
1719 static void iomap_writepage_end_bio(struct bio *bio)
1720 {
1721 	iomap_finish_ioend(iomap_ioend_from_bio(bio),
1722 			blk_status_to_errno(bio->bi_status));
1723 }
1724 
1725 /*
1726  * Submit the final bio for an ioend.
1727  *
1728  * If @error is non-zero, it means that we have a situation where some part of
1729  * the submission process has failed after we've marked pages for writeback.
1730  * We cannot cancel ioend directly in that case, so call the bio end I/O handler
1731  * with the error status here to run the normal I/O completion handler to clear
1732  * the writeback bit and let the file system proess the errors.
1733  */
1734 static int iomap_submit_ioend(struct iomap_writepage_ctx *wpc, int error)
1735 {
1736 	if (!wpc->ioend)
1737 		return error;
1738 
1739 	/*
1740 	 * Let the file systems prepare the I/O submission and hook in an I/O
1741 	 * comletion handler.  This also needs to happen in case after a
1742 	 * failure happened so that the file system end I/O handler gets called
1743 	 * to clean up.
1744 	 */
1745 	if (wpc->ops->prepare_ioend)
1746 		error = wpc->ops->prepare_ioend(wpc->ioend, error);
1747 
1748 	if (error) {
1749 		wpc->ioend->io_bio.bi_status = errno_to_blk_status(error);
1750 		bio_endio(&wpc->ioend->io_bio);
1751 	} else {
1752 		submit_bio(&wpc->ioend->io_bio);
1753 	}
1754 
1755 	wpc->ioend = NULL;
1756 	return error;
1757 }
1758 
1759 static struct iomap_ioend *iomap_alloc_ioend(struct iomap_writepage_ctx *wpc,
1760 		struct writeback_control *wbc, struct inode *inode, loff_t pos)
1761 {
1762 	struct iomap_ioend *ioend;
1763 	struct bio *bio;
1764 
1765 	bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1766 			       REQ_OP_WRITE | wbc_to_write_flags(wbc),
1767 			       GFP_NOFS, &iomap_ioend_bioset);
1768 	bio->bi_iter.bi_sector = iomap_sector(&wpc->iomap, pos);
1769 	bio->bi_end_io = iomap_writepage_end_bio;
1770 	wbc_init_bio(wbc, bio);
1771 	bio->bi_write_hint = inode->i_write_hint;
1772 
1773 	ioend = iomap_ioend_from_bio(bio);
1774 	INIT_LIST_HEAD(&ioend->io_list);
1775 	ioend->io_type = wpc->iomap.type;
1776 	ioend->io_flags = wpc->iomap.flags;
1777 	ioend->io_inode = inode;
1778 	ioend->io_size = 0;
1779 	ioend->io_offset = pos;
1780 	ioend->io_sector = bio->bi_iter.bi_sector;
1781 
1782 	wpc->nr_folios = 0;
1783 	return ioend;
1784 }
1785 
1786 static bool iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t pos)
1787 {
1788 	if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1789 	    (wpc->ioend->io_flags & IOMAP_F_SHARED))
1790 		return false;
1791 	if (wpc->iomap.type != wpc->ioend->io_type)
1792 		return false;
1793 	if (pos != wpc->ioend->io_offset + wpc->ioend->io_size)
1794 		return false;
1795 	if (iomap_sector(&wpc->iomap, pos) !=
1796 	    bio_end_sector(&wpc->ioend->io_bio))
1797 		return false;
1798 	/*
1799 	 * Limit ioend bio chain lengths to minimise IO completion latency. This
1800 	 * also prevents long tight loops ending page writeback on all the
1801 	 * folios in the ioend.
1802 	 */
1803 	if (wpc->nr_folios >= IOEND_BATCH_SIZE)
1804 		return false;
1805 	return true;
1806 }
1807 
1808 /*
1809  * Test to see if we have an existing ioend structure that we could append to
1810  * first; otherwise finish off the current ioend and start another.
1811  *
1812  * If a new ioend is created and cached, the old ioend is submitted to the block
1813  * layer instantly.  Batching optimisations are provided by higher level block
1814  * plugging.
1815  *
1816  * At the end of a writeback pass, there will be a cached ioend remaining on the
1817  * writepage context that the caller will need to submit.
1818  */
1819 static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc,
1820 		struct writeback_control *wbc, struct folio *folio,
1821 		struct inode *inode, loff_t pos, unsigned len)
1822 {
1823 	struct iomap_folio_state *ifs = folio->private;
1824 	size_t poff = offset_in_folio(folio, pos);
1825 	int error;
1826 
1827 	if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos)) {
1828 new_ioend:
1829 		error = iomap_submit_ioend(wpc, 0);
1830 		if (error)
1831 			return error;
1832 		wpc->ioend = iomap_alloc_ioend(wpc, wbc, inode, pos);
1833 	}
1834 
1835 	if (!bio_add_folio(&wpc->ioend->io_bio, folio, len, poff))
1836 		goto new_ioend;
1837 
1838 	if (ifs)
1839 		atomic_add(len, &ifs->write_bytes_pending);
1840 	wpc->ioend->io_size += len;
1841 	wbc_account_cgroup_owner(wbc, &folio->page, len);
1842 	return 0;
1843 }
1844 
1845 static int iomap_writepage_map_blocks(struct iomap_writepage_ctx *wpc,
1846 		struct writeback_control *wbc, struct folio *folio,
1847 		struct inode *inode, u64 pos, unsigned dirty_len,
1848 		unsigned *count)
1849 {
1850 	int error;
1851 
1852 	do {
1853 		unsigned map_len;
1854 
1855 		error = wpc->ops->map_blocks(wpc, inode, pos, dirty_len);
1856 		if (error)
1857 			break;
1858 		trace_iomap_writepage_map(inode, pos, dirty_len, &wpc->iomap);
1859 
1860 		map_len = min_t(u64, dirty_len,
1861 			wpc->iomap.offset + wpc->iomap.length - pos);
1862 		WARN_ON_ONCE(!folio->private && map_len < dirty_len);
1863 
1864 		switch (wpc->iomap.type) {
1865 		case IOMAP_INLINE:
1866 			WARN_ON_ONCE(1);
1867 			error = -EIO;
1868 			break;
1869 		case IOMAP_HOLE:
1870 			break;
1871 		default:
1872 			error = iomap_add_to_ioend(wpc, wbc, folio, inode, pos,
1873 					map_len);
1874 			if (!error)
1875 				(*count)++;
1876 			break;
1877 		}
1878 		dirty_len -= map_len;
1879 		pos += map_len;
1880 	} while (dirty_len && !error);
1881 
1882 	/*
1883 	 * We cannot cancel the ioend directly here on error.  We may have
1884 	 * already set other pages under writeback and hence we have to run I/O
1885 	 * completion to mark the error state of the pages under writeback
1886 	 * appropriately.
1887 	 *
1888 	 * Just let the file system know what portion of the folio failed to
1889 	 * map.
1890 	 */
1891 	if (error && wpc->ops->discard_folio)
1892 		wpc->ops->discard_folio(folio, pos);
1893 	return error;
1894 }
1895 
1896 /*
1897  * Check interaction of the folio with the file end.
1898  *
1899  * If the folio is entirely beyond i_size, return false.  If it straddles
1900  * i_size, adjust end_pos and zero all data beyond i_size.
1901  */
1902 static bool iomap_writepage_handle_eof(struct folio *folio, struct inode *inode,
1903 		u64 *end_pos)
1904 {
1905 	u64 isize = i_size_read(inode);
1906 
1907 	if (*end_pos > isize) {
1908 		size_t poff = offset_in_folio(folio, isize);
1909 		pgoff_t end_index = isize >> PAGE_SHIFT;
1910 
1911 		/*
1912 		 * If the folio is entirely ouside of i_size, skip it.
1913 		 *
1914 		 * This can happen due to a truncate operation that is in
1915 		 * progress and in that case truncate will finish it off once
1916 		 * we've dropped the folio lock.
1917 		 *
1918 		 * Note that the pgoff_t used for end_index is an unsigned long.
1919 		 * If the given offset is greater than 16TB on a 32-bit system,
1920 		 * then if we checked if the folio is fully outside i_size with
1921 		 * "if (folio->index >= end_index + 1)", "end_index + 1" would
1922 		 * overflow and evaluate to 0.  Hence this folio would be
1923 		 * redirtied and written out repeatedly, which would result in
1924 		 * an infinite loop; the user program performing this operation
1925 		 * would hang.  Instead, we can detect this situation by
1926 		 * checking if the folio is totally beyond i_size or if its
1927 		 * offset is just equal to the EOF.
1928 		 */
1929 		if (folio->index > end_index ||
1930 		    (folio->index == end_index && poff == 0))
1931 			return false;
1932 
1933 		/*
1934 		 * The folio straddles i_size.
1935 		 *
1936 		 * It must be zeroed out on each and every writepage invocation
1937 		 * because it may be mmapped:
1938 		 *
1939 		 *    A file is mapped in multiples of the page size.  For a
1940 		 *    file that is not a multiple of the page size, the
1941 		 *    remaining memory is zeroed when mapped, and writes to that
1942 		 *    region are not written out to the file.
1943 		 *
1944 		 * Also adjust the writeback range to skip all blocks entirely
1945 		 * beyond i_size.
1946 		 */
1947 		folio_zero_segment(folio, poff, folio_size(folio));
1948 		*end_pos = round_up(isize, i_blocksize(inode));
1949 	}
1950 
1951 	return true;
1952 }
1953 
1954 static int iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1955 		struct writeback_control *wbc, struct folio *folio)
1956 {
1957 	struct iomap_folio_state *ifs = folio->private;
1958 	struct inode *inode = folio->mapping->host;
1959 	u64 pos = folio_pos(folio);
1960 	u64 end_pos = pos + folio_size(folio);
1961 	unsigned count = 0;
1962 	int error = 0;
1963 	u32 rlen;
1964 
1965 	WARN_ON_ONCE(!folio_test_locked(folio));
1966 	WARN_ON_ONCE(folio_test_dirty(folio));
1967 	WARN_ON_ONCE(folio_test_writeback(folio));
1968 
1969 	trace_iomap_writepage(inode, pos, folio_size(folio));
1970 
1971 	if (!iomap_writepage_handle_eof(folio, inode, &end_pos)) {
1972 		folio_unlock(folio);
1973 		return 0;
1974 	}
1975 	WARN_ON_ONCE(end_pos <= pos);
1976 
1977 	if (i_blocks_per_folio(inode, folio) > 1) {
1978 		if (!ifs) {
1979 			ifs = ifs_alloc(inode, folio, 0);
1980 			iomap_set_range_dirty(folio, 0, end_pos - pos);
1981 		}
1982 
1983 		/*
1984 		 * Keep the I/O completion handler from clearing the writeback
1985 		 * bit until we have submitted all blocks by adding a bias to
1986 		 * ifs->write_bytes_pending, which is dropped after submitting
1987 		 * all blocks.
1988 		 */
1989 		WARN_ON_ONCE(atomic_read(&ifs->write_bytes_pending) != 0);
1990 		atomic_inc(&ifs->write_bytes_pending);
1991 	}
1992 
1993 	/*
1994 	 * Set the writeback bit ASAP, as the I/O completion for the single
1995 	 * block per folio case happen hit as soon as we're submitting the bio.
1996 	 */
1997 	folio_start_writeback(folio);
1998 
1999 	/*
2000 	 * Walk through the folio to find dirty areas to write back.
2001 	 */
2002 	while ((rlen = iomap_find_dirty_range(folio, &pos, end_pos))) {
2003 		error = iomap_writepage_map_blocks(wpc, wbc, folio, inode,
2004 				pos, rlen, &count);
2005 		if (error)
2006 			break;
2007 		pos += rlen;
2008 	}
2009 
2010 	if (count)
2011 		wpc->nr_folios++;
2012 
2013 	/*
2014 	 * We can have dirty bits set past end of file in page_mkwrite path
2015 	 * while mapping the last partial folio. Hence it's better to clear
2016 	 * all the dirty bits in the folio here.
2017 	 */
2018 	iomap_clear_range_dirty(folio, 0, folio_size(folio));
2019 
2020 	/*
2021 	 * Usually the writeback bit is cleared by the I/O completion handler.
2022 	 * But we may end up either not actually writing any blocks, or (when
2023 	 * there are multiple blocks in a folio) all I/O might have finished
2024 	 * already at this point.  In that case we need to clear the writeback
2025 	 * bit ourselves right after unlocking the page.
2026 	 */
2027 	folio_unlock(folio);
2028 	if (ifs) {
2029 		if (atomic_dec_and_test(&ifs->write_bytes_pending))
2030 			folio_end_writeback(folio);
2031 	} else {
2032 		if (!count)
2033 			folio_end_writeback(folio);
2034 	}
2035 	mapping_set_error(inode->i_mapping, error);
2036 	return error;
2037 }
2038 
2039 int
2040 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
2041 		struct iomap_writepage_ctx *wpc,
2042 		const struct iomap_writeback_ops *ops)
2043 {
2044 	struct folio *folio = NULL;
2045 	int error;
2046 
2047 	/*
2048 	 * Writeback from reclaim context should never happen except in the case
2049 	 * of a VM regression so warn about it and refuse to write the data.
2050 	 */
2051 	if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC | PF_KSWAPD)) ==
2052 			PF_MEMALLOC))
2053 		return -EIO;
2054 
2055 	wpc->ops = ops;
2056 	while ((folio = writeback_iter(mapping, wbc, folio, &error)))
2057 		error = iomap_writepage_map(wpc, wbc, folio);
2058 	return iomap_submit_ioend(wpc, error);
2059 }
2060 EXPORT_SYMBOL_GPL(iomap_writepages);
2061 
2062 static int __init iomap_buffered_init(void)
2063 {
2064 	return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
2065 			   offsetof(struct iomap_ioend, io_bio),
2066 			   BIOSET_NEED_BVECS);
2067 }
2068 fs_initcall(iomap_buffered_init);
2069