1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (C) 2016-2019 Christoph Hellwig.
6 #include <linux/module.h>
7 #include <linux/compiler.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>
22 #include "../internal.h"
24 #define IOEND_BATCH_SIZE 4096
26 typedef int (*iomap_punch_t)(struct inode *inode, loff_t offset, loff_t length);
28 * Structure allocated for each folio to track per-block uptodate, dirty state
29 * and I/O completions.
31 struct iomap_folio_state {
32 spinlock_t state_lock;
33 unsigned int read_bytes_pending;
34 atomic_t write_bytes_pending;
37 * Each block has two bits in this bitmap:
38 * Bits [0..blocks_per_folio) has the uptodate status.
39 * Bits [b_p_f...(2*b_p_f)) has the dirty status.
41 unsigned long state[];
44 static struct bio_set iomap_ioend_bioset;
46 static inline bool ifs_is_fully_uptodate(struct folio *folio,
47 struct iomap_folio_state *ifs)
49 struct inode *inode = folio->mapping->host;
51 return bitmap_full(ifs->state, i_blocks_per_folio(inode, folio));
54 static inline bool ifs_block_is_uptodate(struct iomap_folio_state *ifs,
57 return test_bit(block, ifs->state);
60 static bool ifs_set_range_uptodate(struct folio *folio,
61 struct iomap_folio_state *ifs, size_t off, size_t len)
63 struct inode *inode = folio->mapping->host;
64 unsigned int first_blk = off >> inode->i_blkbits;
65 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
66 unsigned int nr_blks = last_blk - first_blk + 1;
68 bitmap_set(ifs->state, first_blk, nr_blks);
69 return ifs_is_fully_uptodate(folio, ifs);
72 static void iomap_set_range_uptodate(struct folio *folio, size_t off,
75 struct iomap_folio_state *ifs = folio->private;
80 spin_lock_irqsave(&ifs->state_lock, flags);
81 uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
82 spin_unlock_irqrestore(&ifs->state_lock, flags);
86 folio_mark_uptodate(folio);
89 static inline bool ifs_block_is_dirty(struct folio *folio,
90 struct iomap_folio_state *ifs, int block)
92 struct inode *inode = folio->mapping->host;
93 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
95 return test_bit(block + blks_per_folio, ifs->state);
98 static void ifs_clear_range_dirty(struct folio *folio,
99 struct iomap_folio_state *ifs, size_t off, size_t len)
101 struct inode *inode = folio->mapping->host;
102 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
103 unsigned int first_blk = (off >> inode->i_blkbits);
104 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
105 unsigned int nr_blks = last_blk - first_blk + 1;
108 spin_lock_irqsave(&ifs->state_lock, flags);
109 bitmap_clear(ifs->state, first_blk + blks_per_folio, nr_blks);
110 spin_unlock_irqrestore(&ifs->state_lock, flags);
113 static void iomap_clear_range_dirty(struct folio *folio, size_t off, size_t len)
115 struct iomap_folio_state *ifs = folio->private;
118 ifs_clear_range_dirty(folio, ifs, off, len);
121 static void ifs_set_range_dirty(struct folio *folio,
122 struct iomap_folio_state *ifs, size_t off, size_t len)
124 struct inode *inode = folio->mapping->host;
125 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
126 unsigned int first_blk = (off >> inode->i_blkbits);
127 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
128 unsigned int nr_blks = last_blk - first_blk + 1;
131 spin_lock_irqsave(&ifs->state_lock, flags);
132 bitmap_set(ifs->state, first_blk + blks_per_folio, nr_blks);
133 spin_unlock_irqrestore(&ifs->state_lock, flags);
136 static void iomap_set_range_dirty(struct folio *folio, size_t off, size_t len)
138 struct iomap_folio_state *ifs = folio->private;
141 ifs_set_range_dirty(folio, ifs, off, len);
144 static struct iomap_folio_state *ifs_alloc(struct inode *inode,
145 struct folio *folio, unsigned int flags)
147 struct iomap_folio_state *ifs = folio->private;
148 unsigned int nr_blocks = i_blocks_per_folio(inode, folio);
151 if (ifs || nr_blocks <= 1)
154 if (flags & IOMAP_NOWAIT)
157 gfp = GFP_NOFS | __GFP_NOFAIL;
160 * ifs->state tracks two sets of state flags when the
161 * filesystem block size is smaller than the folio size.
162 * The first state tracks per-block uptodate and the
163 * second tracks per-block dirty state.
165 ifs = kzalloc(struct_size(ifs, state,
166 BITS_TO_LONGS(2 * nr_blocks)), gfp);
170 spin_lock_init(&ifs->state_lock);
171 if (folio_test_uptodate(folio))
172 bitmap_set(ifs->state, 0, nr_blocks);
173 if (folio_test_dirty(folio))
174 bitmap_set(ifs->state, nr_blocks, nr_blocks);
175 folio_attach_private(folio, ifs);
180 static void ifs_free(struct folio *folio)
182 struct iomap_folio_state *ifs = folio_detach_private(folio);
186 WARN_ON_ONCE(ifs->read_bytes_pending != 0);
187 WARN_ON_ONCE(atomic_read(&ifs->write_bytes_pending));
188 WARN_ON_ONCE(ifs_is_fully_uptodate(folio, ifs) !=
189 folio_test_uptodate(folio));
194 * Calculate the range inside the folio that we actually need to read.
196 static void iomap_adjust_read_range(struct inode *inode, struct folio *folio,
197 loff_t *pos, loff_t length, size_t *offp, size_t *lenp)
199 struct iomap_folio_state *ifs = folio->private;
200 loff_t orig_pos = *pos;
201 loff_t isize = i_size_read(inode);
202 unsigned block_bits = inode->i_blkbits;
203 unsigned block_size = (1 << block_bits);
204 size_t poff = offset_in_folio(folio, *pos);
205 size_t plen = min_t(loff_t, folio_size(folio) - poff, length);
206 unsigned first = poff >> block_bits;
207 unsigned last = (poff + plen - 1) >> block_bits;
210 * If the block size is smaller than the page size, we need to check the
211 * per-block uptodate status and adjust the offset and length if needed
212 * to avoid reading in already uptodate ranges.
217 /* move forward for each leading block marked uptodate */
218 for (i = first; i <= last; i++) {
219 if (!ifs_block_is_uptodate(ifs, i))
227 /* truncate len if we find any trailing uptodate block(s) */
228 for ( ; i <= last; i++) {
229 if (ifs_block_is_uptodate(ifs, i)) {
230 plen -= (last - i + 1) * block_size;
238 * If the extent spans the block that contains the i_size, we need to
239 * handle both halves separately so that we properly zero data in the
240 * page cache for blocks that are entirely outside of i_size.
242 if (orig_pos <= isize && orig_pos + length > isize) {
243 unsigned end = offset_in_folio(folio, isize - 1) >> block_bits;
245 if (first <= end && last > end)
246 plen -= (last - end) * block_size;
253 static void iomap_finish_folio_read(struct folio *folio, size_t off,
254 size_t len, int error)
256 struct iomap_folio_state *ifs = folio->private;
257 bool uptodate = !error;
258 bool finished = true;
263 spin_lock_irqsave(&ifs->state_lock, flags);
265 uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
266 ifs->read_bytes_pending -= len;
267 finished = !ifs->read_bytes_pending;
268 spin_unlock_irqrestore(&ifs->state_lock, flags);
272 folio_set_error(folio);
274 folio_end_read(folio, uptodate);
277 static void iomap_read_end_io(struct bio *bio)
279 int error = blk_status_to_errno(bio->bi_status);
280 struct folio_iter fi;
282 bio_for_each_folio_all(fi, bio)
283 iomap_finish_folio_read(fi.folio, fi.offset, fi.length, error);
287 struct iomap_readpage_ctx {
288 struct folio *cur_folio;
289 bool cur_folio_in_bio;
291 struct readahead_control *rac;
295 * iomap_read_inline_data - copy inline data into the page cache
296 * @iter: iteration structure
297 * @folio: folio to copy to
299 * Copy the inline data in @iter into @folio and zero out the rest of the folio.
300 * Only a single IOMAP_INLINE extent is allowed at the end of each file.
301 * Returns zero for success to complete the read, or the usual negative errno.
303 static int iomap_read_inline_data(const struct iomap_iter *iter,
306 const struct iomap *iomap = iomap_iter_srcmap(iter);
307 size_t size = i_size_read(iter->inode) - iomap->offset;
308 size_t offset = offset_in_folio(folio, iomap->offset);
310 if (folio_test_uptodate(folio))
313 if (WARN_ON_ONCE(size > iomap->length))
316 ifs_alloc(iter->inode, folio, iter->flags);
318 folio_fill_tail(folio, offset, iomap->inline_data, size);
319 iomap_set_range_uptodate(folio, offset, folio_size(folio) - offset);
323 static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
326 const struct iomap *srcmap = iomap_iter_srcmap(iter);
328 return srcmap->type != IOMAP_MAPPED ||
329 (srcmap->flags & IOMAP_F_NEW) ||
330 pos >= i_size_read(iter->inode);
333 static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
334 struct iomap_readpage_ctx *ctx, loff_t offset)
336 const struct iomap *iomap = &iter->iomap;
337 loff_t pos = iter->pos + offset;
338 loff_t length = iomap_length(iter) - offset;
339 struct folio *folio = ctx->cur_folio;
340 struct iomap_folio_state *ifs;
341 loff_t orig_pos = pos;
345 if (iomap->type == IOMAP_INLINE)
346 return iomap_read_inline_data(iter, folio);
348 /* zero post-eof blocks as the page may be mapped */
349 ifs = ifs_alloc(iter->inode, folio, iter->flags);
350 iomap_adjust_read_range(iter->inode, folio, &pos, length, &poff, &plen);
354 if (iomap_block_needs_zeroing(iter, pos)) {
355 folio_zero_range(folio, poff, plen);
356 iomap_set_range_uptodate(folio, poff, plen);
360 ctx->cur_folio_in_bio = true;
362 spin_lock_irq(&ifs->state_lock);
363 ifs->read_bytes_pending += plen;
364 spin_unlock_irq(&ifs->state_lock);
367 sector = iomap_sector(iomap, pos);
369 bio_end_sector(ctx->bio) != sector ||
370 !bio_add_folio(ctx->bio, folio, plen, poff)) {
371 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
372 gfp_t orig_gfp = gfp;
373 unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
376 submit_bio(ctx->bio);
378 if (ctx->rac) /* same as readahead_gfp_mask */
379 gfp |= __GFP_NORETRY | __GFP_NOWARN;
380 ctx->bio = bio_alloc(iomap->bdev, bio_max_segs(nr_vecs),
383 * If the bio_alloc fails, try it again for a single page to
384 * avoid having to deal with partial page reads. This emulates
385 * what do_mpage_read_folio does.
388 ctx->bio = bio_alloc(iomap->bdev, 1, REQ_OP_READ,
392 ctx->bio->bi_opf |= REQ_RAHEAD;
393 ctx->bio->bi_iter.bi_sector = sector;
394 ctx->bio->bi_end_io = iomap_read_end_io;
395 bio_add_folio_nofail(ctx->bio, folio, plen, poff);
400 * Move the caller beyond our range so that it keeps making progress.
401 * For that, we have to include any leading non-uptodate ranges, but
402 * we can skip trailing ones as they will be handled in the next
405 return pos - orig_pos + plen;
408 int iomap_read_folio(struct folio *folio, const struct iomap_ops *ops)
410 struct iomap_iter iter = {
411 .inode = folio->mapping->host,
412 .pos = folio_pos(folio),
413 .len = folio_size(folio),
415 struct iomap_readpage_ctx ctx = {
420 trace_iomap_readpage(iter.inode, 1);
422 while ((ret = iomap_iter(&iter, ops)) > 0)
423 iter.processed = iomap_readpage_iter(&iter, &ctx, 0);
426 folio_set_error(folio);
430 WARN_ON_ONCE(!ctx.cur_folio_in_bio);
432 WARN_ON_ONCE(ctx.cur_folio_in_bio);
437 * Just like mpage_readahead and block_read_full_folio, we always
438 * return 0 and just set the folio error flag on errors. This
439 * should be cleaned up throughout the stack eventually.
443 EXPORT_SYMBOL_GPL(iomap_read_folio);
445 static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
446 struct iomap_readpage_ctx *ctx)
448 loff_t length = iomap_length(iter);
451 for (done = 0; done < length; done += ret) {
452 if (ctx->cur_folio &&
453 offset_in_folio(ctx->cur_folio, iter->pos + done) == 0) {
454 if (!ctx->cur_folio_in_bio)
455 folio_unlock(ctx->cur_folio);
456 ctx->cur_folio = NULL;
458 if (!ctx->cur_folio) {
459 ctx->cur_folio = readahead_folio(ctx->rac);
460 ctx->cur_folio_in_bio = false;
462 ret = iomap_readpage_iter(iter, ctx, done);
471 * iomap_readahead - Attempt to read pages from a file.
472 * @rac: Describes the pages to be read.
473 * @ops: The operations vector for the filesystem.
475 * This function is for filesystems to call to implement their readahead
476 * address_space operation.
478 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
479 * blocks from disc), and may wait for it. The caller may be trying to
480 * access a different page, and so sleeping excessively should be avoided.
481 * It may allocate memory, but should avoid costly allocations. This
482 * function is called with memalloc_nofs set, so allocations will not cause
483 * the filesystem to be reentered.
485 void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
487 struct iomap_iter iter = {
488 .inode = rac->mapping->host,
489 .pos = readahead_pos(rac),
490 .len = readahead_length(rac),
492 struct iomap_readpage_ctx ctx = {
496 trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
498 while (iomap_iter(&iter, ops) > 0)
499 iter.processed = iomap_readahead_iter(&iter, &ctx);
504 if (!ctx.cur_folio_in_bio)
505 folio_unlock(ctx.cur_folio);
508 EXPORT_SYMBOL_GPL(iomap_readahead);
511 * iomap_is_partially_uptodate checks whether blocks within a folio are
514 * Returns true if all blocks which correspond to the specified part
515 * of the folio are uptodate.
517 bool iomap_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
519 struct iomap_folio_state *ifs = folio->private;
520 struct inode *inode = folio->mapping->host;
521 unsigned first, last, i;
526 /* Caller's range may extend past the end of this folio */
527 count = min(folio_size(folio) - from, count);
529 /* First and last blocks in range within folio */
530 first = from >> inode->i_blkbits;
531 last = (from + count - 1) >> inode->i_blkbits;
533 for (i = first; i <= last; i++)
534 if (!ifs_block_is_uptodate(ifs, i))
538 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
541 * iomap_get_folio - get a folio reference for writing
542 * @iter: iteration structure
543 * @pos: start offset of write
544 * @len: Suggested size of folio to create.
546 * Returns a locked reference to the folio at @pos, or an error pointer if the
547 * folio could not be obtained.
549 struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
551 fgf_t fgp = FGP_WRITEBEGIN | FGP_NOFS;
553 if (iter->flags & IOMAP_NOWAIT)
555 fgp |= fgf_set_order(len);
557 return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
558 fgp, mapping_gfp_mask(iter->inode->i_mapping));
560 EXPORT_SYMBOL_GPL(iomap_get_folio);
562 bool iomap_release_folio(struct folio *folio, gfp_t gfp_flags)
564 trace_iomap_release_folio(folio->mapping->host, folio_pos(folio),
568 * If the folio is dirty, we refuse to release our metadata because
569 * it may be partially dirty. Once we track per-block dirty state,
570 * we can release the metadata if every block is dirty.
572 if (folio_test_dirty(folio))
577 EXPORT_SYMBOL_GPL(iomap_release_folio);
579 void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
581 trace_iomap_invalidate_folio(folio->mapping->host,
582 folio_pos(folio) + offset, len);
585 * If we're invalidating the entire folio, clear the dirty state
586 * from it and release it to avoid unnecessary buildup of the LRU.
588 if (offset == 0 && len == folio_size(folio)) {
589 WARN_ON_ONCE(folio_test_writeback(folio));
590 folio_cancel_dirty(folio);
594 EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
596 bool iomap_dirty_folio(struct address_space *mapping, struct folio *folio)
598 struct inode *inode = mapping->host;
599 size_t len = folio_size(folio);
601 ifs_alloc(inode, folio, 0);
602 iomap_set_range_dirty(folio, 0, len);
603 return filemap_dirty_folio(mapping, folio);
605 EXPORT_SYMBOL_GPL(iomap_dirty_folio);
608 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
610 loff_t i_size = i_size_read(inode);
613 * Only truncate newly allocated pages beyoned EOF, even if the
614 * write started inside the existing inode size.
616 if (pos + len > i_size)
617 truncate_pagecache_range(inode, max(pos, i_size),
621 static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
622 size_t poff, size_t plen, const struct iomap *iomap)
627 bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
628 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
629 bio_add_folio_nofail(&bio, folio, plen, poff);
630 return submit_bio_wait(&bio);
633 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
634 size_t len, struct folio *folio)
636 const struct iomap *srcmap = iomap_iter_srcmap(iter);
637 struct iomap_folio_state *ifs;
638 loff_t block_size = i_blocksize(iter->inode);
639 loff_t block_start = round_down(pos, block_size);
640 loff_t block_end = round_up(pos + len, block_size);
641 unsigned int nr_blocks = i_blocks_per_folio(iter->inode, folio);
642 size_t from = offset_in_folio(folio, pos), to = from + len;
646 * If the write or zeroing completely overlaps the current folio, then
647 * entire folio will be dirtied so there is no need for
648 * per-block state tracking structures to be attached to this folio.
649 * For the unshare case, we must read in the ondisk contents because we
650 * are not changing pagecache contents.
652 if (!(iter->flags & IOMAP_UNSHARE) && pos <= folio_pos(folio) &&
653 pos + len >= folio_pos(folio) + folio_size(folio))
656 ifs = ifs_alloc(iter->inode, folio, iter->flags);
657 if ((iter->flags & IOMAP_NOWAIT) && !ifs && nr_blocks > 1)
660 if (folio_test_uptodate(folio))
662 folio_clear_error(folio);
665 iomap_adjust_read_range(iter->inode, folio, &block_start,
666 block_end - block_start, &poff, &plen);
670 if (!(iter->flags & IOMAP_UNSHARE) &&
671 (from <= poff || from >= poff + plen) &&
672 (to <= poff || to >= poff + plen))
675 if (iomap_block_needs_zeroing(iter, block_start)) {
676 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
678 folio_zero_segments(folio, poff, from, to, poff + plen);
682 if (iter->flags & IOMAP_NOWAIT)
685 status = iomap_read_folio_sync(block_start, folio,
690 iomap_set_range_uptodate(folio, poff, plen);
691 } while ((block_start += plen) < block_end);
696 static struct folio *__iomap_get_folio(struct iomap_iter *iter, loff_t pos,
699 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
701 if (folio_ops && folio_ops->get_folio)
702 return folio_ops->get_folio(iter, pos, len);
704 return iomap_get_folio(iter, pos, len);
707 static void __iomap_put_folio(struct iomap_iter *iter, loff_t pos, size_t ret,
710 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
712 if (folio_ops && folio_ops->put_folio) {
713 folio_ops->put_folio(iter->inode, pos, ret, folio);
720 static int iomap_write_begin_inline(const struct iomap_iter *iter,
723 /* needs more work for the tailpacking case; disable for now */
724 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
726 return iomap_read_inline_data(iter, folio);
729 static int iomap_write_begin(struct iomap_iter *iter, loff_t pos,
730 size_t len, struct folio **foliop)
732 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
733 const struct iomap *srcmap = iomap_iter_srcmap(iter);
737 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
738 if (srcmap != &iter->iomap)
739 BUG_ON(pos + len > srcmap->offset + srcmap->length);
741 if (fatal_signal_pending(current))
744 if (!mapping_large_folio_support(iter->inode->i_mapping))
745 len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
747 folio = __iomap_get_folio(iter, pos, len);
749 return PTR_ERR(folio);
752 * Now we have a locked folio, before we do anything with it we need to
753 * check that the iomap we have cached is not stale. The inode extent
754 * mapping can change due to concurrent IO in flight (e.g.
755 * IOMAP_UNWRITTEN state can change and memory reclaim could have
756 * reclaimed a previously partially written page at this index after IO
757 * completion before this write reaches this file offset) and hence we
758 * could do the wrong thing here (zero a page range incorrectly or fail
759 * to zero) and corrupt data.
761 if (folio_ops && folio_ops->iomap_valid) {
762 bool iomap_valid = folio_ops->iomap_valid(iter->inode,
765 iter->iomap.flags |= IOMAP_F_STALE;
771 if (pos + len > folio_pos(folio) + folio_size(folio))
772 len = folio_pos(folio) + folio_size(folio) - pos;
774 if (srcmap->type == IOMAP_INLINE)
775 status = iomap_write_begin_inline(iter, folio);
776 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
777 status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
779 status = __iomap_write_begin(iter, pos, len, folio);
781 if (unlikely(status))
788 __iomap_put_folio(iter, pos, 0, folio);
789 iomap_write_failed(iter->inode, pos, len);
794 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
795 size_t copied, struct folio *folio)
797 flush_dcache_folio(folio);
800 * The blocks that were entirely written will now be uptodate, so we
801 * don't have to worry about a read_folio reading them and overwriting a
802 * partial write. However, if we've encountered a short write and only
803 * partially written into a block, it will not be marked uptodate, so a
804 * read_folio might come in and destroy our partial write.
806 * Do the simplest thing and just treat any short write to a
807 * non-uptodate page as a zero-length write, and force the caller to
808 * redo the whole thing.
810 if (unlikely(copied < len && !folio_test_uptodate(folio)))
812 iomap_set_range_uptodate(folio, offset_in_folio(folio, pos), len);
813 iomap_set_range_dirty(folio, offset_in_folio(folio, pos), copied);
814 filemap_dirty_folio(inode->i_mapping, folio);
818 static size_t iomap_write_end_inline(const struct iomap_iter *iter,
819 struct folio *folio, loff_t pos, size_t copied)
821 const struct iomap *iomap = &iter->iomap;
824 WARN_ON_ONCE(!folio_test_uptodate(folio));
825 BUG_ON(!iomap_inline_data_valid(iomap));
827 flush_dcache_folio(folio);
828 addr = kmap_local_folio(folio, pos);
829 memcpy(iomap_inline_data(iomap, pos), addr, copied);
832 mark_inode_dirty(iter->inode);
836 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
837 static size_t iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
838 size_t copied, struct folio *folio)
840 const struct iomap *srcmap = iomap_iter_srcmap(iter);
841 loff_t old_size = iter->inode->i_size;
844 if (srcmap->type == IOMAP_INLINE) {
845 ret = iomap_write_end_inline(iter, folio, pos, copied);
846 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
847 ret = block_write_end(NULL, iter->inode->i_mapping, pos, len,
848 copied, &folio->page, NULL);
850 ret = __iomap_write_end(iter->inode, pos, len, copied, folio);
854 * Update the in-memory inode size after copying the data into the page
855 * cache. It's up to the file system to write the updated size to disk,
856 * preferably after I/O completion so that no stale data is exposed.
858 if (pos + ret > old_size) {
859 i_size_write(iter->inode, pos + ret);
860 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
862 __iomap_put_folio(iter, pos, ret, folio);
865 pagecache_isize_extended(iter->inode, old_size, pos);
867 iomap_write_failed(iter->inode, pos + ret, len - ret);
871 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
873 loff_t length = iomap_length(iter);
874 size_t chunk = PAGE_SIZE << MAX_PAGECACHE_ORDER;
875 loff_t pos = iter->pos;
878 struct address_space *mapping = iter->inode->i_mapping;
879 unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
883 size_t offset; /* Offset into folio */
884 size_t bytes; /* Bytes to write to folio */
885 size_t copied; /* Bytes copied from user */
887 bytes = iov_iter_count(i);
889 offset = pos & (chunk - 1);
890 bytes = min(chunk - offset, bytes);
891 status = balance_dirty_pages_ratelimited_flags(mapping,
893 if (unlikely(status))
900 * Bring in the user page that we'll copy from _first_.
901 * Otherwise there's a nasty deadlock on copying from the
902 * same page as we're writing to, without it being marked
905 * For async buffered writes the assumption is that the user
906 * page has already been faulted in. This can be optimized by
907 * faulting the user page.
909 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
914 status = iomap_write_begin(iter, pos, bytes, &folio);
915 if (unlikely(status))
917 if (iter->iomap.flags & IOMAP_F_STALE)
920 offset = offset_in_folio(folio, pos);
921 if (bytes > folio_size(folio) - offset)
922 bytes = folio_size(folio) - offset;
924 if (mapping_writably_mapped(mapping))
925 flush_dcache_folio(folio);
927 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
928 status = iomap_write_end(iter, pos, bytes, copied, folio);
930 if (unlikely(copied != status))
931 iov_iter_revert(i, copied - status);
934 if (unlikely(status == 0)) {
936 * A short copy made iomap_write_end() reject the
937 * thing entirely. Might be memory poisoning
938 * halfway through, might be a race with munmap,
939 * might be severe memory pressure.
941 if (chunk > PAGE_SIZE)
952 } while (iov_iter_count(i) && length);
954 if (status == -EAGAIN) {
955 iov_iter_revert(i, written);
958 return written ? written : status;
962 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
963 const struct iomap_ops *ops)
965 struct iomap_iter iter = {
966 .inode = iocb->ki_filp->f_mapping->host,
968 .len = iov_iter_count(i),
969 .flags = IOMAP_WRITE,
973 if (iocb->ki_flags & IOCB_NOWAIT)
974 iter.flags |= IOMAP_NOWAIT;
976 while ((ret = iomap_iter(&iter, ops)) > 0)
977 iter.processed = iomap_write_iter(&iter, i);
979 if (unlikely(iter.pos == iocb->ki_pos))
981 ret = iter.pos - iocb->ki_pos;
982 iocb->ki_pos = iter.pos;
985 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
987 static int iomap_write_delalloc_ifs_punch(struct inode *inode,
988 struct folio *folio, loff_t start_byte, loff_t end_byte,
991 unsigned int first_blk, last_blk, i;
993 u8 blkbits = inode->i_blkbits;
994 struct iomap_folio_state *ifs;
998 * When we have per-block dirty tracking, there can be
999 * blocks within a folio which are marked uptodate
1000 * but not dirty. In that case it is necessary to punch
1001 * out such blocks to avoid leaking any delalloc blocks.
1003 ifs = folio->private;
1007 last_byte = min_t(loff_t, end_byte - 1,
1008 folio_pos(folio) + folio_size(folio) - 1);
1009 first_blk = offset_in_folio(folio, start_byte) >> blkbits;
1010 last_blk = offset_in_folio(folio, last_byte) >> blkbits;
1011 for (i = first_blk; i <= last_blk; i++) {
1012 if (!ifs_block_is_dirty(folio, ifs, i)) {
1013 ret = punch(inode, folio_pos(folio) + (i << blkbits),
1024 static int iomap_write_delalloc_punch(struct inode *inode, struct folio *folio,
1025 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1026 iomap_punch_t punch)
1030 if (!folio_test_dirty(folio))
1033 /* if dirty, punch up to offset */
1034 if (start_byte > *punch_start_byte) {
1035 ret = punch(inode, *punch_start_byte,
1036 start_byte - *punch_start_byte);
1041 /* Punch non-dirty blocks within folio */
1042 ret = iomap_write_delalloc_ifs_punch(inode, folio, start_byte,
1048 * Make sure the next punch start is correctly bound to
1049 * the end of this data range, not the end of the folio.
1051 *punch_start_byte = min_t(loff_t, end_byte,
1052 folio_pos(folio) + folio_size(folio));
1058 * Scan the data range passed to us for dirty page cache folios. If we find a
1059 * dirty folio, punch out the preceding range and update the offset from which
1060 * the next punch will start from.
1062 * We can punch out storage reservations under clean pages because they either
1063 * contain data that has been written back - in which case the delalloc punch
1064 * over that range is a no-op - or they have been read faults in which case they
1065 * contain zeroes and we can remove the delalloc backing range and any new
1066 * writes to those pages will do the normal hole filling operation...
1068 * This makes the logic simple: we only need to keep the delalloc extents only
1069 * over the dirty ranges of the page cache.
1071 * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1072 * simplify range iterations.
1074 static int iomap_write_delalloc_scan(struct inode *inode,
1075 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1076 iomap_punch_t punch)
1078 while (start_byte < end_byte) {
1079 struct folio *folio;
1082 /* grab locked page */
1083 folio = filemap_lock_folio(inode->i_mapping,
1084 start_byte >> PAGE_SHIFT);
1085 if (IS_ERR(folio)) {
1086 start_byte = ALIGN_DOWN(start_byte, PAGE_SIZE) +
1091 ret = iomap_write_delalloc_punch(inode, folio, punch_start_byte,
1092 start_byte, end_byte, punch);
1094 folio_unlock(folio);
1099 /* move offset to start of next folio in range */
1100 start_byte = folio_next_index(folio) << PAGE_SHIFT;
1101 folio_unlock(folio);
1108 * Punch out all the delalloc blocks in the range given except for those that
1109 * have dirty data still pending in the page cache - those are going to be
1110 * written and so must still retain the delalloc backing for writeback.
1112 * As we are scanning the page cache for data, we don't need to reimplement the
1113 * wheel - mapping_seek_hole_data() does exactly what we need to identify the
1114 * start and end of data ranges correctly even for sub-folio block sizes. This
1115 * byte range based iteration is especially convenient because it means we
1116 * don't have to care about variable size folios, nor where the start or end of
1117 * the data range lies within a folio, if they lie within the same folio or even
1118 * if there are multiple discontiguous data ranges within the folio.
1120 * It should be noted that mapping_seek_hole_data() is not aware of EOF, and so
1121 * can return data ranges that exist in the cache beyond EOF. e.g. a page fault
1122 * spanning EOF will initialise the post-EOF data to zeroes and mark it up to
1123 * date. A write page fault can then mark it dirty. If we then fail a write()
1124 * beyond EOF into that up to date cached range, we allocate a delalloc block
1125 * beyond EOF and then have to punch it out. Because the range is up to date,
1126 * mapping_seek_hole_data() will return it, and we will skip the punch because
1127 * the folio is dirty. THis is incorrect - we always need to punch out delalloc
1128 * beyond EOF in this case as writeback will never write back and covert that
1129 * delalloc block beyond EOF. Hence we limit the cached data scan range to EOF,
1130 * resulting in always punching out the range from the EOF to the end of the
1131 * range the iomap spans.
1133 * Intervals are of the form [start_byte, end_byte) (i.e. open ended) because it
1134 * matches the intervals returned by mapping_seek_hole_data(). i.e. SEEK_DATA
1135 * returns the start of a data range (start_byte), and SEEK_HOLE(start_byte)
1136 * returns the end of the data range (data_end). Using closed intervals would
1137 * require sprinkling this code with magic "+ 1" and "- 1" arithmetic and expose
1138 * the code to subtle off-by-one bugs....
1140 static int iomap_write_delalloc_release(struct inode *inode,
1141 loff_t start_byte, loff_t end_byte, iomap_punch_t punch)
1143 loff_t punch_start_byte = start_byte;
1144 loff_t scan_end_byte = min(i_size_read(inode), end_byte);
1148 * Lock the mapping to avoid races with page faults re-instantiating
1149 * folios and dirtying them via ->page_mkwrite whilst we walk the
1150 * cache and perform delalloc extent removal. Failing to do this can
1151 * leave dirty pages with no space reservation in the cache.
1153 filemap_invalidate_lock(inode->i_mapping);
1154 while (start_byte < scan_end_byte) {
1157 start_byte = mapping_seek_hole_data(inode->i_mapping,
1158 start_byte, scan_end_byte, SEEK_DATA);
1160 * If there is no more data to scan, all that is left is to
1161 * punch out the remaining range.
1163 if (start_byte == -ENXIO || start_byte == scan_end_byte)
1165 if (start_byte < 0) {
1169 WARN_ON_ONCE(start_byte < punch_start_byte);
1170 WARN_ON_ONCE(start_byte > scan_end_byte);
1173 * We find the end of this contiguous cached data range by
1174 * seeking from start_byte to the beginning of the next hole.
1176 data_end = mapping_seek_hole_data(inode->i_mapping, start_byte,
1177 scan_end_byte, SEEK_HOLE);
1182 WARN_ON_ONCE(data_end <= start_byte);
1183 WARN_ON_ONCE(data_end > scan_end_byte);
1185 error = iomap_write_delalloc_scan(inode, &punch_start_byte,
1186 start_byte, data_end, punch);
1190 /* The next data search starts at the end of this one. */
1191 start_byte = data_end;
1194 if (punch_start_byte < end_byte)
1195 error = punch(inode, punch_start_byte,
1196 end_byte - punch_start_byte);
1198 filemap_invalidate_unlock(inode->i_mapping);
1203 * When a short write occurs, the filesystem may need to remove reserved space
1204 * that was allocated in ->iomap_begin from it's ->iomap_end method. For
1205 * filesystems that use delayed allocation, we need to punch out delalloc
1206 * extents from the range that are not dirty in the page cache. As the write can
1207 * race with page faults, there can be dirty pages over the delalloc extent
1208 * outside the range of a short write but still within the delalloc extent
1209 * allocated for this iomap.
1211 * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1212 * simplify range iterations.
1214 * The punch() callback *must* only punch delalloc extents in the range passed
1215 * to it. It must skip over all other types of extents in the range and leave
1216 * them completely unchanged. It must do this punch atomically with respect to
1217 * other extent modifications.
1219 * The punch() callback may be called with a folio locked to prevent writeback
1220 * extent allocation racing at the edge of the range we are currently punching.
1221 * The locked folio may or may not cover the range being punched, so it is not
1222 * safe for the punch() callback to lock folios itself.
1226 * inode->i_rwsem (shared or exclusive)
1227 * inode->i_mapping->invalidate_lock (exclusive)
1230 * internal filesystem allocation lock
1232 int iomap_file_buffered_write_punch_delalloc(struct inode *inode,
1233 struct iomap *iomap, loff_t pos, loff_t length,
1234 ssize_t written, iomap_punch_t punch)
1238 unsigned int blocksize = i_blocksize(inode);
1240 if (iomap->type != IOMAP_DELALLOC)
1243 /* If we didn't reserve the blocks, we're not allowed to punch them. */
1244 if (!(iomap->flags & IOMAP_F_NEW))
1248 * start_byte refers to the first unused block after a short write. If
1249 * nothing was written, round offset down to point at the first block in
1252 if (unlikely(!written))
1253 start_byte = round_down(pos, blocksize);
1255 start_byte = round_up(pos + written, blocksize);
1256 end_byte = round_up(pos + length, blocksize);
1258 /* Nothing to do if we've written the entire delalloc extent */
1259 if (start_byte >= end_byte)
1262 return iomap_write_delalloc_release(inode, start_byte, end_byte,
1265 EXPORT_SYMBOL_GPL(iomap_file_buffered_write_punch_delalloc);
1267 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
1269 struct iomap *iomap = &iter->iomap;
1270 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1271 loff_t pos = iter->pos;
1272 loff_t length = iomap_length(iter);
1275 /* don't bother with blocks that are not shared to start with */
1276 if (!(iomap->flags & IOMAP_F_SHARED))
1278 /* don't bother with holes or unwritten extents */
1279 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1283 struct folio *folio;
1286 size_t bytes = min_t(u64, SIZE_MAX, length);
1288 status = iomap_write_begin(iter, pos, bytes, &folio);
1289 if (unlikely(status))
1291 if (iomap->flags & IOMAP_F_STALE)
1294 offset = offset_in_folio(folio, pos);
1295 if (bytes > folio_size(folio) - offset)
1296 bytes = folio_size(folio) - offset;
1298 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
1299 if (WARN_ON_ONCE(bytes == 0))
1308 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
1309 } while (length > 0);
1315 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1316 const struct iomap_ops *ops)
1318 struct iomap_iter iter = {
1322 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
1326 while ((ret = iomap_iter(&iter, ops)) > 0)
1327 iter.processed = iomap_unshare_iter(&iter);
1330 EXPORT_SYMBOL_GPL(iomap_file_unshare);
1332 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
1334 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1335 loff_t pos = iter->pos;
1336 loff_t length = iomap_length(iter);
1339 /* already zeroed? we're done. */
1340 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1344 struct folio *folio;
1347 size_t bytes = min_t(u64, SIZE_MAX, length);
1349 status = iomap_write_begin(iter, pos, bytes, &folio);
1352 if (iter->iomap.flags & IOMAP_F_STALE)
1355 offset = offset_in_folio(folio, pos);
1356 if (bytes > folio_size(folio) - offset)
1357 bytes = folio_size(folio) - offset;
1359 folio_zero_range(folio, offset, bytes);
1360 folio_mark_accessed(folio);
1362 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
1363 if (WARN_ON_ONCE(bytes == 0))
1369 } while (length > 0);
1377 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1378 const struct iomap_ops *ops)
1380 struct iomap_iter iter = {
1384 .flags = IOMAP_ZERO,
1388 while ((ret = iomap_iter(&iter, ops)) > 0)
1389 iter.processed = iomap_zero_iter(&iter, did_zero);
1392 EXPORT_SYMBOL_GPL(iomap_zero_range);
1395 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1396 const struct iomap_ops *ops)
1398 unsigned int blocksize = i_blocksize(inode);
1399 unsigned int off = pos & (blocksize - 1);
1401 /* Block boundary? Nothing to do */
1404 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1406 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1408 static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
1409 struct folio *folio)
1411 loff_t length = iomap_length(iter);
1414 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
1415 ret = __block_write_begin_int(folio, iter->pos, length, NULL,
1419 block_commit_write(&folio->page, 0, length);
1421 WARN_ON_ONCE(!folio_test_uptodate(folio));
1422 folio_mark_dirty(folio);
1428 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1430 struct iomap_iter iter = {
1431 .inode = file_inode(vmf->vma->vm_file),
1432 .flags = IOMAP_WRITE | IOMAP_FAULT,
1434 struct folio *folio = page_folio(vmf->page);
1438 ret = folio_mkwrite_check_truncate(folio, iter.inode);
1441 iter.pos = folio_pos(folio);
1443 while ((ret = iomap_iter(&iter, ops)) > 0)
1444 iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1448 folio_wait_stable(folio);
1449 return VM_FAULT_LOCKED;
1451 folio_unlock(folio);
1452 return vmf_fs_error(ret);
1454 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1456 static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1457 size_t len, int error)
1459 struct iomap_folio_state *ifs = folio->private;
1462 folio_set_error(folio);
1463 mapping_set_error(inode->i_mapping, error);
1466 WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !ifs);
1467 WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) <= 0);
1469 if (!ifs || atomic_sub_and_test(len, &ifs->write_bytes_pending))
1470 folio_end_writeback(folio);
1474 * We're now finished for good with this ioend structure. Update the page
1475 * state, release holds on bios, and finally free up memory. Do not use the
1479 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1481 struct inode *inode = ioend->io_inode;
1482 struct bio *bio = &ioend->io_inline_bio;
1483 struct bio *last = ioend->io_bio, *next;
1484 u64 start = bio->bi_iter.bi_sector;
1485 loff_t offset = ioend->io_offset;
1486 bool quiet = bio_flagged(bio, BIO_QUIET);
1487 u32 folio_count = 0;
1489 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1490 struct folio_iter fi;
1493 * For the last bio, bi_private points to the ioend, so we
1494 * need to explicitly end the iteration here.
1499 next = bio->bi_private;
1501 /* walk all folios in bio, ending page IO on them */
1502 bio_for_each_folio_all(fi, bio) {
1503 iomap_finish_folio_write(inode, fi.folio, fi.length,
1509 /* The ioend has been freed by bio_put() */
1511 if (unlikely(error && !quiet)) {
1512 printk_ratelimited(KERN_ERR
1513 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1514 inode->i_sb->s_id, inode->i_ino, offset, start);
1520 * Ioend completion routine for merged bios. This can only be called from task
1521 * contexts as merged ioends can be of unbound length. Hence we have to break up
1522 * the writeback completions into manageable chunks to avoid long scheduler
1523 * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1524 * good batch processing throughput without creating adverse scheduler latency
1528 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1530 struct list_head tmp;
1535 list_replace_init(&ioend->io_list, &tmp);
1536 completions = iomap_finish_ioend(ioend, error);
1538 while (!list_empty(&tmp)) {
1539 if (completions > IOEND_BATCH_SIZE * 8) {
1543 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1544 list_del_init(&ioend->io_list);
1545 completions += iomap_finish_ioend(ioend, error);
1548 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1551 * We can merge two adjacent ioends if they have the same set of work to do.
1554 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1556 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1558 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1559 (next->io_flags & IOMAP_F_SHARED))
1561 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1562 (next->io_type == IOMAP_UNWRITTEN))
1564 if (ioend->io_offset + ioend->io_size != next->io_offset)
1567 * Do not merge physically discontiguous ioends. The filesystem
1568 * completion functions will have to iterate the physical
1569 * discontiguities even if we merge the ioends at a logical level, so
1570 * we don't gain anything by merging physical discontiguities here.
1572 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1573 * submission so does not point to the start sector of the bio at
1576 if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1582 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1584 struct iomap_ioend *next;
1586 INIT_LIST_HEAD(&ioend->io_list);
1588 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1590 if (!iomap_ioend_can_merge(ioend, next))
1592 list_move_tail(&next->io_list, &ioend->io_list);
1593 ioend->io_size += next->io_size;
1596 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1599 iomap_ioend_compare(void *priv, const struct list_head *a,
1600 const struct list_head *b)
1602 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1603 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1605 if (ia->io_offset < ib->io_offset)
1607 if (ia->io_offset > ib->io_offset)
1613 iomap_sort_ioends(struct list_head *ioend_list)
1615 list_sort(NULL, ioend_list, iomap_ioend_compare);
1617 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1619 static void iomap_writepage_end_bio(struct bio *bio)
1621 struct iomap_ioend *ioend = bio->bi_private;
1623 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1627 * Submit the final bio for an ioend.
1629 * If @error is non-zero, it means that we have a situation where some part of
1630 * the submission process has failed after we've marked pages for writeback
1631 * and unlocked them. In this situation, we need to fail the bio instead of
1632 * submitting it. This typically only happens on a filesystem shutdown.
1635 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1638 ioend->io_bio->bi_private = ioend;
1639 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1641 if (wpc->ops->prepare_ioend)
1642 error = wpc->ops->prepare_ioend(ioend, error);
1645 * If we're failing the IO now, just mark the ioend with an
1646 * error and finish it. This will run IO completion immediately
1647 * as there is only one reference to the ioend at this point in
1650 ioend->io_bio->bi_status = errno_to_blk_status(error);
1651 bio_endio(ioend->io_bio);
1655 submit_bio(ioend->io_bio);
1659 static struct iomap_ioend *
1660 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1661 loff_t offset, sector_t sector, struct writeback_control *wbc)
1663 struct iomap_ioend *ioend;
1666 bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1667 REQ_OP_WRITE | wbc_to_write_flags(wbc),
1668 GFP_NOFS, &iomap_ioend_bioset);
1669 bio->bi_iter.bi_sector = sector;
1670 wbc_init_bio(wbc, bio);
1672 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1673 INIT_LIST_HEAD(&ioend->io_list);
1674 ioend->io_type = wpc->iomap.type;
1675 ioend->io_flags = wpc->iomap.flags;
1676 ioend->io_inode = inode;
1678 ioend->io_folios = 0;
1679 ioend->io_offset = offset;
1680 ioend->io_bio = bio;
1681 ioend->io_sector = sector;
1686 * Allocate a new bio, and chain the old bio to the new one.
1688 * Note that we have to perform the chaining in this unintuitive order
1689 * so that the bi_private linkage is set up in the right direction for the
1690 * traversal in iomap_finish_ioend().
1693 iomap_chain_bio(struct bio *prev)
1697 new = bio_alloc(prev->bi_bdev, BIO_MAX_VECS, prev->bi_opf, GFP_NOFS);
1698 bio_clone_blkg_association(new, prev);
1699 new->bi_iter.bi_sector = bio_end_sector(prev);
1701 bio_chain(prev, new);
1702 bio_get(prev); /* for iomap_finish_ioend */
1708 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1711 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1712 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1714 if (wpc->iomap.type != wpc->ioend->io_type)
1716 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1718 if (sector != bio_end_sector(wpc->ioend->io_bio))
1721 * Limit ioend bio chain lengths to minimise IO completion latency. This
1722 * also prevents long tight loops ending page writeback on all the
1723 * folios in the ioend.
1725 if (wpc->ioend->io_folios >= IOEND_BATCH_SIZE)
1731 * Test to see if we have an existing ioend structure that we could append to
1732 * first; otherwise finish off the current ioend and start another.
1735 iomap_add_to_ioend(struct inode *inode, loff_t pos, struct folio *folio,
1736 struct iomap_folio_state *ifs, struct iomap_writepage_ctx *wpc,
1737 struct writeback_control *wbc, struct list_head *iolist)
1739 sector_t sector = iomap_sector(&wpc->iomap, pos);
1740 unsigned len = i_blocksize(inode);
1741 size_t poff = offset_in_folio(folio, pos);
1743 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos, sector)) {
1745 list_add(&wpc->ioend->io_list, iolist);
1746 wpc->ioend = iomap_alloc_ioend(inode, wpc, pos, sector, wbc);
1749 if (!bio_add_folio(wpc->ioend->io_bio, folio, len, poff)) {
1750 wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1751 bio_add_folio_nofail(wpc->ioend->io_bio, folio, len, poff);
1755 atomic_add(len, &ifs->write_bytes_pending);
1756 wpc->ioend->io_size += len;
1757 wbc_account_cgroup_owner(wbc, &folio->page, len);
1761 * We implement an immediate ioend submission policy here to avoid needing to
1762 * chain multiple ioends and hence nest mempool allocations which can violate
1763 * the forward progress guarantees we need to provide. The current ioend we're
1764 * adding blocks to is cached in the writepage context, and if the new block
1765 * doesn't append to the cached ioend, it will create a new ioend and cache that
1768 * If a new ioend is created and cached, the old ioend is returned and queued
1769 * locally for submission once the entire page is processed or an error has been
1770 * detected. While ioends are submitted immediately after they are completed,
1771 * batching optimisations are provided by higher level block plugging.
1773 * At the end of a writeback pass, there will be a cached ioend remaining on the
1774 * writepage context that the caller will need to submit.
1777 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1778 struct writeback_control *wbc, struct inode *inode,
1779 struct folio *folio, u64 end_pos)
1781 struct iomap_folio_state *ifs = folio->private;
1782 struct iomap_ioend *ioend, *next;
1783 unsigned len = i_blocksize(inode);
1784 unsigned nblocks = i_blocks_per_folio(inode, folio);
1785 u64 pos = folio_pos(folio);
1786 int error = 0, count = 0, i;
1787 LIST_HEAD(submit_list);
1789 WARN_ON_ONCE(end_pos <= pos);
1791 if (!ifs && nblocks > 1) {
1792 ifs = ifs_alloc(inode, folio, 0);
1793 iomap_set_range_dirty(folio, 0, end_pos - pos);
1796 WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) != 0);
1799 * Walk through the folio to find areas to write back. If we
1800 * run off the end of the current map or find the current map
1801 * invalid, grab a new one.
1803 for (i = 0; i < nblocks && pos < end_pos; i++, pos += len) {
1804 if (ifs && !ifs_block_is_dirty(folio, ifs, i))
1807 error = wpc->ops->map_blocks(wpc, inode, pos);
1810 trace_iomap_writepage_map(inode, &wpc->iomap);
1811 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1813 if (wpc->iomap.type == IOMAP_HOLE)
1815 iomap_add_to_ioend(inode, pos, folio, ifs, wpc, wbc,
1820 wpc->ioend->io_folios++;
1822 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1823 WARN_ON_ONCE(!folio_test_locked(folio));
1824 WARN_ON_ONCE(folio_test_writeback(folio));
1825 WARN_ON_ONCE(folio_test_dirty(folio));
1828 * We cannot cancel the ioend directly here on error. We may have
1829 * already set other pages under writeback and hence we have to run I/O
1830 * completion to mark the error state of the pages under writeback
1833 if (unlikely(error)) {
1835 * Let the filesystem know what portion of the current page
1836 * failed to map. If the page hasn't been added to ioend, it
1837 * won't be affected by I/O completion and we must unlock it
1840 if (wpc->ops->discard_folio)
1841 wpc->ops->discard_folio(folio, pos);
1843 folio_unlock(folio);
1849 * We can have dirty bits set past end of file in page_mkwrite path
1850 * while mapping the last partial folio. Hence it's better to clear
1851 * all the dirty bits in the folio here.
1853 iomap_clear_range_dirty(folio, 0, folio_size(folio));
1854 folio_start_writeback(folio);
1855 folio_unlock(folio);
1858 * Preserve the original error if there was one; catch
1859 * submission errors here and propagate into subsequent ioend
1862 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1865 list_del_init(&ioend->io_list);
1866 error2 = iomap_submit_ioend(wpc, ioend, error);
1867 if (error2 && !error)
1872 * We can end up here with no error and nothing to write only if we race
1873 * with a partial page truncate on a sub-page block sized filesystem.
1876 folio_end_writeback(folio);
1878 mapping_set_error(inode->i_mapping, error);
1883 * Write out a dirty page.
1885 * For delalloc space on the page, we need to allocate space and flush it.
1886 * For unwritten space on the page, we need to start the conversion to
1887 * regular allocated space.
1889 static int iomap_do_writepage(struct folio *folio,
1890 struct writeback_control *wbc, void *data)
1892 struct iomap_writepage_ctx *wpc = data;
1893 struct inode *inode = folio->mapping->host;
1896 trace_iomap_writepage(inode, folio_pos(folio), folio_size(folio));
1899 * Refuse to write the folio out if we're called from reclaim context.
1901 * This avoids stack overflows when called from deeply used stacks in
1902 * random callers for direct reclaim or memcg reclaim. We explicitly
1903 * allow reclaim from kswapd as the stack usage there is relatively low.
1905 * This should never happen except in the case of a VM regression so
1908 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1913 * Is this folio beyond the end of the file?
1915 * The folio index is less than the end_index, adjust the end_pos
1916 * to the highest offset that this folio should represent.
1917 * -----------------------------------------------------
1918 * | file mapping | <EOF> |
1919 * -----------------------------------------------------
1920 * | Page ... | Page N-2 | Page N-1 | Page N | |
1921 * ^--------------------------------^----------|--------
1922 * | desired writeback range | see else |
1923 * ---------------------------------^------------------|
1925 isize = i_size_read(inode);
1926 end_pos = folio_pos(folio) + folio_size(folio);
1927 if (end_pos > isize) {
1929 * Check whether the page to write out is beyond or straddles
1931 * -------------------------------------------------------
1932 * | file mapping | <EOF> |
1933 * -------------------------------------------------------
1934 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1935 * ^--------------------------------^-----------|---------
1937 * ---------------------------------^-----------|--------|
1939 size_t poff = offset_in_folio(folio, isize);
1940 pgoff_t end_index = isize >> PAGE_SHIFT;
1943 * Skip the page if it's fully outside i_size, e.g.
1944 * due to a truncate operation that's in progress. We've
1945 * cleaned this page and truncate will finish things off for
1948 * Note that the end_index is unsigned long. If the given
1949 * offset is greater than 16TB on a 32-bit system then if we
1950 * checked if the page is fully outside i_size with
1951 * "if (page->index >= end_index + 1)", "end_index + 1" would
1952 * overflow and evaluate to 0. Hence this page would be
1953 * redirtied and written out repeatedly, which would result in
1954 * an infinite loop; the user program performing this operation
1955 * would hang. Instead, we can detect this situation by
1956 * checking if the page is totally beyond i_size or if its
1957 * offset is just equal to the EOF.
1959 if (folio->index > end_index ||
1960 (folio->index == end_index && poff == 0))
1964 * The page straddles i_size. It must be zeroed out on each
1965 * and every writepage invocation because it may be mmapped.
1966 * "A file is mapped in multiples of the page size. For a file
1967 * that is not a multiple of the page size, the remaining
1968 * memory is zeroed when mapped, and writes to that region are
1969 * not written out to the file."
1971 folio_zero_segment(folio, poff, folio_size(folio));
1975 return iomap_writepage_map(wpc, wbc, inode, folio, end_pos);
1978 folio_redirty_for_writepage(wbc, folio);
1980 folio_unlock(folio);
1985 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1986 struct iomap_writepage_ctx *wpc,
1987 const struct iomap_writeback_ops *ops)
1992 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1995 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1997 EXPORT_SYMBOL_GPL(iomap_writepages);
1999 static int __init iomap_init(void)
2001 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
2002 offsetof(struct iomap_ioend, io_inline_bio),
2005 fs_initcall(iomap_init);