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 poff = offset_in_page(iomap->offset);
309 size_t offset = offset_in_folio(folio, iomap->offset);
312 if (folio_test_uptodate(folio))
315 if (WARN_ON_ONCE(size > PAGE_SIZE - poff))
317 if (WARN_ON_ONCE(size > PAGE_SIZE -
318 offset_in_page(iomap->inline_data)))
320 if (WARN_ON_ONCE(size > iomap->length))
323 ifs_alloc(iter->inode, folio, iter->flags);
325 addr = kmap_local_folio(folio, offset);
326 memcpy(addr, iomap->inline_data, size);
327 memset(addr + size, 0, PAGE_SIZE - poff - size);
329 iomap_set_range_uptodate(folio, offset, PAGE_SIZE - poff);
333 static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
336 const struct iomap *srcmap = iomap_iter_srcmap(iter);
338 return srcmap->type != IOMAP_MAPPED ||
339 (srcmap->flags & IOMAP_F_NEW) ||
340 pos >= i_size_read(iter->inode);
343 static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
344 struct iomap_readpage_ctx *ctx, loff_t offset)
346 const struct iomap *iomap = &iter->iomap;
347 loff_t pos = iter->pos + offset;
348 loff_t length = iomap_length(iter) - offset;
349 struct folio *folio = ctx->cur_folio;
350 struct iomap_folio_state *ifs;
351 loff_t orig_pos = pos;
355 if (iomap->type == IOMAP_INLINE)
356 return iomap_read_inline_data(iter, folio);
358 /* zero post-eof blocks as the page may be mapped */
359 ifs = ifs_alloc(iter->inode, folio, iter->flags);
360 iomap_adjust_read_range(iter->inode, folio, &pos, length, &poff, &plen);
364 if (iomap_block_needs_zeroing(iter, pos)) {
365 folio_zero_range(folio, poff, plen);
366 iomap_set_range_uptodate(folio, poff, plen);
370 ctx->cur_folio_in_bio = true;
372 spin_lock_irq(&ifs->state_lock);
373 ifs->read_bytes_pending += plen;
374 spin_unlock_irq(&ifs->state_lock);
377 sector = iomap_sector(iomap, pos);
379 bio_end_sector(ctx->bio) != sector ||
380 !bio_add_folio(ctx->bio, folio, plen, poff)) {
381 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
382 gfp_t orig_gfp = gfp;
383 unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
386 submit_bio(ctx->bio);
388 if (ctx->rac) /* same as readahead_gfp_mask */
389 gfp |= __GFP_NORETRY | __GFP_NOWARN;
390 ctx->bio = bio_alloc(iomap->bdev, bio_max_segs(nr_vecs),
393 * If the bio_alloc fails, try it again for a single page to
394 * avoid having to deal with partial page reads. This emulates
395 * what do_mpage_read_folio does.
398 ctx->bio = bio_alloc(iomap->bdev, 1, REQ_OP_READ,
402 ctx->bio->bi_opf |= REQ_RAHEAD;
403 ctx->bio->bi_iter.bi_sector = sector;
404 ctx->bio->bi_end_io = iomap_read_end_io;
405 bio_add_folio_nofail(ctx->bio, folio, plen, poff);
410 * Move the caller beyond our range so that it keeps making progress.
411 * For that, we have to include any leading non-uptodate ranges, but
412 * we can skip trailing ones as they will be handled in the next
415 return pos - orig_pos + plen;
418 int iomap_read_folio(struct folio *folio, const struct iomap_ops *ops)
420 struct iomap_iter iter = {
421 .inode = folio->mapping->host,
422 .pos = folio_pos(folio),
423 .len = folio_size(folio),
425 struct iomap_readpage_ctx ctx = {
430 trace_iomap_readpage(iter.inode, 1);
432 while ((ret = iomap_iter(&iter, ops)) > 0)
433 iter.processed = iomap_readpage_iter(&iter, &ctx, 0);
436 folio_set_error(folio);
440 WARN_ON_ONCE(!ctx.cur_folio_in_bio);
442 WARN_ON_ONCE(ctx.cur_folio_in_bio);
447 * Just like mpage_readahead and block_read_full_folio, we always
448 * return 0 and just set the folio error flag on errors. This
449 * should be cleaned up throughout the stack eventually.
453 EXPORT_SYMBOL_GPL(iomap_read_folio);
455 static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
456 struct iomap_readpage_ctx *ctx)
458 loff_t length = iomap_length(iter);
461 for (done = 0; done < length; done += ret) {
462 if (ctx->cur_folio &&
463 offset_in_folio(ctx->cur_folio, iter->pos + done) == 0) {
464 if (!ctx->cur_folio_in_bio)
465 folio_unlock(ctx->cur_folio);
466 ctx->cur_folio = NULL;
468 if (!ctx->cur_folio) {
469 ctx->cur_folio = readahead_folio(ctx->rac);
470 ctx->cur_folio_in_bio = false;
472 ret = iomap_readpage_iter(iter, ctx, done);
481 * iomap_readahead - Attempt to read pages from a file.
482 * @rac: Describes the pages to be read.
483 * @ops: The operations vector for the filesystem.
485 * This function is for filesystems to call to implement their readahead
486 * address_space operation.
488 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
489 * blocks from disc), and may wait for it. The caller may be trying to
490 * access a different page, and so sleeping excessively should be avoided.
491 * It may allocate memory, but should avoid costly allocations. This
492 * function is called with memalloc_nofs set, so allocations will not cause
493 * the filesystem to be reentered.
495 void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
497 struct iomap_iter iter = {
498 .inode = rac->mapping->host,
499 .pos = readahead_pos(rac),
500 .len = readahead_length(rac),
502 struct iomap_readpage_ctx ctx = {
506 trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
508 while (iomap_iter(&iter, ops) > 0)
509 iter.processed = iomap_readahead_iter(&iter, &ctx);
514 if (!ctx.cur_folio_in_bio)
515 folio_unlock(ctx.cur_folio);
518 EXPORT_SYMBOL_GPL(iomap_readahead);
521 * iomap_is_partially_uptodate checks whether blocks within a folio are
524 * Returns true if all blocks which correspond to the specified part
525 * of the folio are uptodate.
527 bool iomap_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
529 struct iomap_folio_state *ifs = folio->private;
530 struct inode *inode = folio->mapping->host;
531 unsigned first, last, i;
536 /* Caller's range may extend past the end of this folio */
537 count = min(folio_size(folio) - from, count);
539 /* First and last blocks in range within folio */
540 first = from >> inode->i_blkbits;
541 last = (from + count - 1) >> inode->i_blkbits;
543 for (i = first; i <= last; i++)
544 if (!ifs_block_is_uptodate(ifs, i))
548 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
551 * iomap_get_folio - get a folio reference for writing
552 * @iter: iteration structure
553 * @pos: start offset of write
554 * @len: Suggested size of folio to create.
556 * Returns a locked reference to the folio at @pos, or an error pointer if the
557 * folio could not be obtained.
559 struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
561 fgf_t fgp = FGP_WRITEBEGIN | FGP_NOFS;
563 if (iter->flags & IOMAP_NOWAIT)
565 fgp |= fgf_set_order(len);
567 return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
568 fgp, mapping_gfp_mask(iter->inode->i_mapping));
570 EXPORT_SYMBOL_GPL(iomap_get_folio);
572 bool iomap_release_folio(struct folio *folio, gfp_t gfp_flags)
574 trace_iomap_release_folio(folio->mapping->host, folio_pos(folio),
578 * If the folio is dirty, we refuse to release our metadata because
579 * it may be partially dirty. Once we track per-block dirty state,
580 * we can release the metadata if every block is dirty.
582 if (folio_test_dirty(folio))
587 EXPORT_SYMBOL_GPL(iomap_release_folio);
589 void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
591 trace_iomap_invalidate_folio(folio->mapping->host,
592 folio_pos(folio) + offset, len);
595 * If we're invalidating the entire folio, clear the dirty state
596 * from it and release it to avoid unnecessary buildup of the LRU.
598 if (offset == 0 && len == folio_size(folio)) {
599 WARN_ON_ONCE(folio_test_writeback(folio));
600 folio_cancel_dirty(folio);
604 EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
606 bool iomap_dirty_folio(struct address_space *mapping, struct folio *folio)
608 struct inode *inode = mapping->host;
609 size_t len = folio_size(folio);
611 ifs_alloc(inode, folio, 0);
612 iomap_set_range_dirty(folio, 0, len);
613 return filemap_dirty_folio(mapping, folio);
615 EXPORT_SYMBOL_GPL(iomap_dirty_folio);
618 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
620 loff_t i_size = i_size_read(inode);
623 * Only truncate newly allocated pages beyoned EOF, even if the
624 * write started inside the existing inode size.
626 if (pos + len > i_size)
627 truncate_pagecache_range(inode, max(pos, i_size),
631 static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
632 size_t poff, size_t plen, const struct iomap *iomap)
637 bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
638 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
639 bio_add_folio_nofail(&bio, folio, plen, poff);
640 return submit_bio_wait(&bio);
643 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
644 size_t len, struct folio *folio)
646 const struct iomap *srcmap = iomap_iter_srcmap(iter);
647 struct iomap_folio_state *ifs;
648 loff_t block_size = i_blocksize(iter->inode);
649 loff_t block_start = round_down(pos, block_size);
650 loff_t block_end = round_up(pos + len, block_size);
651 unsigned int nr_blocks = i_blocks_per_folio(iter->inode, folio);
652 size_t from = offset_in_folio(folio, pos), to = from + len;
656 * If the write or zeroing completely overlaps the current folio, then
657 * entire folio will be dirtied so there is no need for
658 * per-block state tracking structures to be attached to this folio.
659 * For the unshare case, we must read in the ondisk contents because we
660 * are not changing pagecache contents.
662 if (!(iter->flags & IOMAP_UNSHARE) && pos <= folio_pos(folio) &&
663 pos + len >= folio_pos(folio) + folio_size(folio))
666 ifs = ifs_alloc(iter->inode, folio, iter->flags);
667 if ((iter->flags & IOMAP_NOWAIT) && !ifs && nr_blocks > 1)
670 if (folio_test_uptodate(folio))
672 folio_clear_error(folio);
675 iomap_adjust_read_range(iter->inode, folio, &block_start,
676 block_end - block_start, &poff, &plen);
680 if (!(iter->flags & IOMAP_UNSHARE) &&
681 (from <= poff || from >= poff + plen) &&
682 (to <= poff || to >= poff + plen))
685 if (iomap_block_needs_zeroing(iter, block_start)) {
686 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
688 folio_zero_segments(folio, poff, from, to, poff + plen);
692 if (iter->flags & IOMAP_NOWAIT)
695 status = iomap_read_folio_sync(block_start, folio,
700 iomap_set_range_uptodate(folio, poff, plen);
701 } while ((block_start += plen) < block_end);
706 static struct folio *__iomap_get_folio(struct iomap_iter *iter, loff_t pos,
709 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
711 if (folio_ops && folio_ops->get_folio)
712 return folio_ops->get_folio(iter, pos, len);
714 return iomap_get_folio(iter, pos, len);
717 static void __iomap_put_folio(struct iomap_iter *iter, loff_t pos, size_t ret,
720 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
722 if (folio_ops && folio_ops->put_folio) {
723 folio_ops->put_folio(iter->inode, pos, ret, folio);
730 static int iomap_write_begin_inline(const struct iomap_iter *iter,
733 /* needs more work for the tailpacking case; disable for now */
734 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
736 return iomap_read_inline_data(iter, folio);
739 static int iomap_write_begin(struct iomap_iter *iter, loff_t pos,
740 size_t len, struct folio **foliop)
742 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
743 const struct iomap *srcmap = iomap_iter_srcmap(iter);
747 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
748 if (srcmap != &iter->iomap)
749 BUG_ON(pos + len > srcmap->offset + srcmap->length);
751 if (fatal_signal_pending(current))
754 if (!mapping_large_folio_support(iter->inode->i_mapping))
755 len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
757 folio = __iomap_get_folio(iter, pos, len);
759 return PTR_ERR(folio);
762 * Now we have a locked folio, before we do anything with it we need to
763 * check that the iomap we have cached is not stale. The inode extent
764 * mapping can change due to concurrent IO in flight (e.g.
765 * IOMAP_UNWRITTEN state can change and memory reclaim could have
766 * reclaimed a previously partially written page at this index after IO
767 * completion before this write reaches this file offset) and hence we
768 * could do the wrong thing here (zero a page range incorrectly or fail
769 * to zero) and corrupt data.
771 if (folio_ops && folio_ops->iomap_valid) {
772 bool iomap_valid = folio_ops->iomap_valid(iter->inode,
775 iter->iomap.flags |= IOMAP_F_STALE;
781 if (pos + len > folio_pos(folio) + folio_size(folio))
782 len = folio_pos(folio) + folio_size(folio) - pos;
784 if (srcmap->type == IOMAP_INLINE)
785 status = iomap_write_begin_inline(iter, folio);
786 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
787 status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
789 status = __iomap_write_begin(iter, pos, len, folio);
791 if (unlikely(status))
798 __iomap_put_folio(iter, pos, 0, folio);
799 iomap_write_failed(iter->inode, pos, len);
804 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
805 size_t copied, struct folio *folio)
807 flush_dcache_folio(folio);
810 * The blocks that were entirely written will now be uptodate, so we
811 * don't have to worry about a read_folio reading them and overwriting a
812 * partial write. However, if we've encountered a short write and only
813 * partially written into a block, it will not be marked uptodate, so a
814 * read_folio might come in and destroy our partial write.
816 * Do the simplest thing and just treat any short write to a
817 * non-uptodate page as a zero-length write, and force the caller to
818 * redo the whole thing.
820 if (unlikely(copied < len && !folio_test_uptodate(folio)))
822 iomap_set_range_uptodate(folio, offset_in_folio(folio, pos), len);
823 iomap_set_range_dirty(folio, offset_in_folio(folio, pos), copied);
824 filemap_dirty_folio(inode->i_mapping, folio);
828 static size_t iomap_write_end_inline(const struct iomap_iter *iter,
829 struct folio *folio, loff_t pos, size_t copied)
831 const struct iomap *iomap = &iter->iomap;
834 WARN_ON_ONCE(!folio_test_uptodate(folio));
835 BUG_ON(!iomap_inline_data_valid(iomap));
837 flush_dcache_folio(folio);
838 addr = kmap_local_folio(folio, pos);
839 memcpy(iomap_inline_data(iomap, pos), addr, copied);
842 mark_inode_dirty(iter->inode);
846 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
847 static size_t iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
848 size_t copied, struct folio *folio)
850 const struct iomap *srcmap = iomap_iter_srcmap(iter);
851 loff_t old_size = iter->inode->i_size;
854 if (srcmap->type == IOMAP_INLINE) {
855 ret = iomap_write_end_inline(iter, folio, pos, copied);
856 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
857 ret = block_write_end(NULL, iter->inode->i_mapping, pos, len,
858 copied, &folio->page, NULL);
860 ret = __iomap_write_end(iter->inode, pos, len, copied, folio);
864 * Update the in-memory inode size after copying the data into the page
865 * cache. It's up to the file system to write the updated size to disk,
866 * preferably after I/O completion so that no stale data is exposed.
868 if (pos + ret > old_size) {
869 i_size_write(iter->inode, pos + ret);
870 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
872 __iomap_put_folio(iter, pos, ret, folio);
875 pagecache_isize_extended(iter->inode, old_size, pos);
877 iomap_write_failed(iter->inode, pos + ret, len - ret);
881 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
883 loff_t length = iomap_length(iter);
884 size_t chunk = PAGE_SIZE << MAX_PAGECACHE_ORDER;
885 loff_t pos = iter->pos;
888 struct address_space *mapping = iter->inode->i_mapping;
889 unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
893 size_t offset; /* Offset into folio */
894 size_t bytes; /* Bytes to write to folio */
895 size_t copied; /* Bytes copied from user */
897 bytes = iov_iter_count(i);
899 offset = pos & (chunk - 1);
900 bytes = min(chunk - offset, bytes);
901 status = balance_dirty_pages_ratelimited_flags(mapping,
903 if (unlikely(status))
910 * Bring in the user page that we'll copy from _first_.
911 * Otherwise there's a nasty deadlock on copying from the
912 * same page as we're writing to, without it being marked
915 * For async buffered writes the assumption is that the user
916 * page has already been faulted in. This can be optimized by
917 * faulting the user page.
919 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
924 status = iomap_write_begin(iter, pos, bytes, &folio);
925 if (unlikely(status))
927 if (iter->iomap.flags & IOMAP_F_STALE)
930 offset = offset_in_folio(folio, pos);
931 if (bytes > folio_size(folio) - offset)
932 bytes = folio_size(folio) - offset;
934 if (mapping_writably_mapped(mapping))
935 flush_dcache_folio(folio);
937 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
938 status = iomap_write_end(iter, pos, bytes, copied, folio);
940 if (unlikely(copied != status))
941 iov_iter_revert(i, copied - status);
944 if (unlikely(status == 0)) {
946 * A short copy made iomap_write_end() reject the
947 * thing entirely. Might be memory poisoning
948 * halfway through, might be a race with munmap,
949 * might be severe memory pressure.
951 if (chunk > PAGE_SIZE)
962 } while (iov_iter_count(i) && length);
964 if (status == -EAGAIN) {
965 iov_iter_revert(i, written);
968 return written ? written : status;
972 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
973 const struct iomap_ops *ops)
975 struct iomap_iter iter = {
976 .inode = iocb->ki_filp->f_mapping->host,
978 .len = iov_iter_count(i),
979 .flags = IOMAP_WRITE,
983 if (iocb->ki_flags & IOCB_NOWAIT)
984 iter.flags |= IOMAP_NOWAIT;
986 while ((ret = iomap_iter(&iter, ops)) > 0)
987 iter.processed = iomap_write_iter(&iter, i);
989 if (unlikely(iter.pos == iocb->ki_pos))
991 ret = iter.pos - iocb->ki_pos;
992 iocb->ki_pos = iter.pos;
995 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
997 static int iomap_write_delalloc_ifs_punch(struct inode *inode,
998 struct folio *folio, loff_t start_byte, loff_t end_byte,
1001 unsigned int first_blk, last_blk, i;
1003 u8 blkbits = inode->i_blkbits;
1004 struct iomap_folio_state *ifs;
1008 * When we have per-block dirty tracking, there can be
1009 * blocks within a folio which are marked uptodate
1010 * but not dirty. In that case it is necessary to punch
1011 * out such blocks to avoid leaking any delalloc blocks.
1013 ifs = folio->private;
1017 last_byte = min_t(loff_t, end_byte - 1,
1018 folio_pos(folio) + folio_size(folio) - 1);
1019 first_blk = offset_in_folio(folio, start_byte) >> blkbits;
1020 last_blk = offset_in_folio(folio, last_byte) >> blkbits;
1021 for (i = first_blk; i <= last_blk; i++) {
1022 if (!ifs_block_is_dirty(folio, ifs, i)) {
1023 ret = punch(inode, folio_pos(folio) + (i << blkbits),
1034 static int iomap_write_delalloc_punch(struct inode *inode, struct folio *folio,
1035 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1036 iomap_punch_t punch)
1040 if (!folio_test_dirty(folio))
1043 /* if dirty, punch up to offset */
1044 if (start_byte > *punch_start_byte) {
1045 ret = punch(inode, *punch_start_byte,
1046 start_byte - *punch_start_byte);
1051 /* Punch non-dirty blocks within folio */
1052 ret = iomap_write_delalloc_ifs_punch(inode, folio, start_byte,
1058 * Make sure the next punch start is correctly bound to
1059 * the end of this data range, not the end of the folio.
1061 *punch_start_byte = min_t(loff_t, end_byte,
1062 folio_pos(folio) + folio_size(folio));
1068 * Scan the data range passed to us for dirty page cache folios. If we find a
1069 * dirty folio, punch out the preceding range and update the offset from which
1070 * the next punch will start from.
1072 * We can punch out storage reservations under clean pages because they either
1073 * contain data that has been written back - in which case the delalloc punch
1074 * over that range is a no-op - or they have been read faults in which case they
1075 * contain zeroes and we can remove the delalloc backing range and any new
1076 * writes to those pages will do the normal hole filling operation...
1078 * This makes the logic simple: we only need to keep the delalloc extents only
1079 * over the dirty ranges of the page cache.
1081 * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1082 * simplify range iterations.
1084 static int iomap_write_delalloc_scan(struct inode *inode,
1085 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1086 iomap_punch_t punch)
1088 while (start_byte < end_byte) {
1089 struct folio *folio;
1092 /* grab locked page */
1093 folio = filemap_lock_folio(inode->i_mapping,
1094 start_byte >> PAGE_SHIFT);
1095 if (IS_ERR(folio)) {
1096 start_byte = ALIGN_DOWN(start_byte, PAGE_SIZE) +
1101 ret = iomap_write_delalloc_punch(inode, folio, punch_start_byte,
1102 start_byte, end_byte, punch);
1104 folio_unlock(folio);
1109 /* move offset to start of next folio in range */
1110 start_byte = folio_next_index(folio) << PAGE_SHIFT;
1111 folio_unlock(folio);
1118 * Punch out all the delalloc blocks in the range given except for those that
1119 * have dirty data still pending in the page cache - those are going to be
1120 * written and so must still retain the delalloc backing for writeback.
1122 * As we are scanning the page cache for data, we don't need to reimplement the
1123 * wheel - mapping_seek_hole_data() does exactly what we need to identify the
1124 * start and end of data ranges correctly even for sub-folio block sizes. This
1125 * byte range based iteration is especially convenient because it means we
1126 * don't have to care about variable size folios, nor where the start or end of
1127 * the data range lies within a folio, if they lie within the same folio or even
1128 * if there are multiple discontiguous data ranges within the folio.
1130 * It should be noted that mapping_seek_hole_data() is not aware of EOF, and so
1131 * can return data ranges that exist in the cache beyond EOF. e.g. a page fault
1132 * spanning EOF will initialise the post-EOF data to zeroes and mark it up to
1133 * date. A write page fault can then mark it dirty. If we then fail a write()
1134 * beyond EOF into that up to date cached range, we allocate a delalloc block
1135 * beyond EOF and then have to punch it out. Because the range is up to date,
1136 * mapping_seek_hole_data() will return it, and we will skip the punch because
1137 * the folio is dirty. THis is incorrect - we always need to punch out delalloc
1138 * beyond EOF in this case as writeback will never write back and covert that
1139 * delalloc block beyond EOF. Hence we limit the cached data scan range to EOF,
1140 * resulting in always punching out the range from the EOF to the end of the
1141 * range the iomap spans.
1143 * Intervals are of the form [start_byte, end_byte) (i.e. open ended) because it
1144 * matches the intervals returned by mapping_seek_hole_data(). i.e. SEEK_DATA
1145 * returns the start of a data range (start_byte), and SEEK_HOLE(start_byte)
1146 * returns the end of the data range (data_end). Using closed intervals would
1147 * require sprinkling this code with magic "+ 1" and "- 1" arithmetic and expose
1148 * the code to subtle off-by-one bugs....
1150 static int iomap_write_delalloc_release(struct inode *inode,
1151 loff_t start_byte, loff_t end_byte, iomap_punch_t punch)
1153 loff_t punch_start_byte = start_byte;
1154 loff_t scan_end_byte = min(i_size_read(inode), end_byte);
1158 * Lock the mapping to avoid races with page faults re-instantiating
1159 * folios and dirtying them via ->page_mkwrite whilst we walk the
1160 * cache and perform delalloc extent removal. Failing to do this can
1161 * leave dirty pages with no space reservation in the cache.
1163 filemap_invalidate_lock(inode->i_mapping);
1164 while (start_byte < scan_end_byte) {
1167 start_byte = mapping_seek_hole_data(inode->i_mapping,
1168 start_byte, scan_end_byte, SEEK_DATA);
1170 * If there is no more data to scan, all that is left is to
1171 * punch out the remaining range.
1173 if (start_byte == -ENXIO || start_byte == scan_end_byte)
1175 if (start_byte < 0) {
1179 WARN_ON_ONCE(start_byte < punch_start_byte);
1180 WARN_ON_ONCE(start_byte > scan_end_byte);
1183 * We find the end of this contiguous cached data range by
1184 * seeking from start_byte to the beginning of the next hole.
1186 data_end = mapping_seek_hole_data(inode->i_mapping, start_byte,
1187 scan_end_byte, SEEK_HOLE);
1192 WARN_ON_ONCE(data_end <= start_byte);
1193 WARN_ON_ONCE(data_end > scan_end_byte);
1195 error = iomap_write_delalloc_scan(inode, &punch_start_byte,
1196 start_byte, data_end, punch);
1200 /* The next data search starts at the end of this one. */
1201 start_byte = data_end;
1204 if (punch_start_byte < end_byte)
1205 error = punch(inode, punch_start_byte,
1206 end_byte - punch_start_byte);
1208 filemap_invalidate_unlock(inode->i_mapping);
1213 * When a short write occurs, the filesystem may need to remove reserved space
1214 * that was allocated in ->iomap_begin from it's ->iomap_end method. For
1215 * filesystems that use delayed allocation, we need to punch out delalloc
1216 * extents from the range that are not dirty in the page cache. As the write can
1217 * race with page faults, there can be dirty pages over the delalloc extent
1218 * outside the range of a short write but still within the delalloc extent
1219 * allocated for this iomap.
1221 * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1222 * simplify range iterations.
1224 * The punch() callback *must* only punch delalloc extents in the range passed
1225 * to it. It must skip over all other types of extents in the range and leave
1226 * them completely unchanged. It must do this punch atomically with respect to
1227 * other extent modifications.
1229 * The punch() callback may be called with a folio locked to prevent writeback
1230 * extent allocation racing at the edge of the range we are currently punching.
1231 * The locked folio may or may not cover the range being punched, so it is not
1232 * safe for the punch() callback to lock folios itself.
1236 * inode->i_rwsem (shared or exclusive)
1237 * inode->i_mapping->invalidate_lock (exclusive)
1240 * internal filesystem allocation lock
1242 int iomap_file_buffered_write_punch_delalloc(struct inode *inode,
1243 struct iomap *iomap, loff_t pos, loff_t length,
1244 ssize_t written, iomap_punch_t punch)
1248 unsigned int blocksize = i_blocksize(inode);
1250 if (iomap->type != IOMAP_DELALLOC)
1253 /* If we didn't reserve the blocks, we're not allowed to punch them. */
1254 if (!(iomap->flags & IOMAP_F_NEW))
1258 * start_byte refers to the first unused block after a short write. If
1259 * nothing was written, round offset down to point at the first block in
1262 if (unlikely(!written))
1263 start_byte = round_down(pos, blocksize);
1265 start_byte = round_up(pos + written, blocksize);
1266 end_byte = round_up(pos + length, blocksize);
1268 /* Nothing to do if we've written the entire delalloc extent */
1269 if (start_byte >= end_byte)
1272 return iomap_write_delalloc_release(inode, start_byte, end_byte,
1275 EXPORT_SYMBOL_GPL(iomap_file_buffered_write_punch_delalloc);
1277 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
1279 struct iomap *iomap = &iter->iomap;
1280 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1281 loff_t pos = iter->pos;
1282 loff_t length = iomap_length(iter);
1285 /* don't bother with blocks that are not shared to start with */
1286 if (!(iomap->flags & IOMAP_F_SHARED))
1288 /* don't bother with holes or unwritten extents */
1289 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1293 struct folio *folio;
1296 size_t bytes = min_t(u64, SIZE_MAX, length);
1298 status = iomap_write_begin(iter, pos, bytes, &folio);
1299 if (unlikely(status))
1301 if (iomap->flags & IOMAP_F_STALE)
1304 offset = offset_in_folio(folio, pos);
1305 if (bytes > folio_size(folio) - offset)
1306 bytes = folio_size(folio) - offset;
1308 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
1309 if (WARN_ON_ONCE(bytes == 0))
1318 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
1319 } while (length > 0);
1325 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1326 const struct iomap_ops *ops)
1328 struct iomap_iter iter = {
1332 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
1336 while ((ret = iomap_iter(&iter, ops)) > 0)
1337 iter.processed = iomap_unshare_iter(&iter);
1340 EXPORT_SYMBOL_GPL(iomap_file_unshare);
1342 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
1344 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1345 loff_t pos = iter->pos;
1346 loff_t length = iomap_length(iter);
1349 /* already zeroed? we're done. */
1350 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1354 struct folio *folio;
1357 size_t bytes = min_t(u64, SIZE_MAX, length);
1359 status = iomap_write_begin(iter, pos, bytes, &folio);
1362 if (iter->iomap.flags & IOMAP_F_STALE)
1365 offset = offset_in_folio(folio, pos);
1366 if (bytes > folio_size(folio) - offset)
1367 bytes = folio_size(folio) - offset;
1369 folio_zero_range(folio, offset, bytes);
1370 folio_mark_accessed(folio);
1372 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
1373 if (WARN_ON_ONCE(bytes == 0))
1379 } while (length > 0);
1387 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1388 const struct iomap_ops *ops)
1390 struct iomap_iter iter = {
1394 .flags = IOMAP_ZERO,
1398 while ((ret = iomap_iter(&iter, ops)) > 0)
1399 iter.processed = iomap_zero_iter(&iter, did_zero);
1402 EXPORT_SYMBOL_GPL(iomap_zero_range);
1405 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1406 const struct iomap_ops *ops)
1408 unsigned int blocksize = i_blocksize(inode);
1409 unsigned int off = pos & (blocksize - 1);
1411 /* Block boundary? Nothing to do */
1414 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1416 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1418 static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
1419 struct folio *folio)
1421 loff_t length = iomap_length(iter);
1424 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
1425 ret = __block_write_begin_int(folio, iter->pos, length, NULL,
1429 block_commit_write(&folio->page, 0, length);
1431 WARN_ON_ONCE(!folio_test_uptodate(folio));
1432 folio_mark_dirty(folio);
1438 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1440 struct iomap_iter iter = {
1441 .inode = file_inode(vmf->vma->vm_file),
1442 .flags = IOMAP_WRITE | IOMAP_FAULT,
1444 struct folio *folio = page_folio(vmf->page);
1448 ret = folio_mkwrite_check_truncate(folio, iter.inode);
1451 iter.pos = folio_pos(folio);
1453 while ((ret = iomap_iter(&iter, ops)) > 0)
1454 iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1458 folio_wait_stable(folio);
1459 return VM_FAULT_LOCKED;
1461 folio_unlock(folio);
1462 return vmf_fs_error(ret);
1464 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1466 static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1467 size_t len, int error)
1469 struct iomap_folio_state *ifs = folio->private;
1472 folio_set_error(folio);
1473 mapping_set_error(inode->i_mapping, error);
1476 WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !ifs);
1477 WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) <= 0);
1479 if (!ifs || atomic_sub_and_test(len, &ifs->write_bytes_pending))
1480 folio_end_writeback(folio);
1484 * We're now finished for good with this ioend structure. Update the page
1485 * state, release holds on bios, and finally free up memory. Do not use the
1489 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1491 struct inode *inode = ioend->io_inode;
1492 struct bio *bio = &ioend->io_inline_bio;
1493 struct bio *last = ioend->io_bio, *next;
1494 u64 start = bio->bi_iter.bi_sector;
1495 loff_t offset = ioend->io_offset;
1496 bool quiet = bio_flagged(bio, BIO_QUIET);
1497 u32 folio_count = 0;
1499 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1500 struct folio_iter fi;
1503 * For the last bio, bi_private points to the ioend, so we
1504 * need to explicitly end the iteration here.
1509 next = bio->bi_private;
1511 /* walk all folios in bio, ending page IO on them */
1512 bio_for_each_folio_all(fi, bio) {
1513 iomap_finish_folio_write(inode, fi.folio, fi.length,
1519 /* The ioend has been freed by bio_put() */
1521 if (unlikely(error && !quiet)) {
1522 printk_ratelimited(KERN_ERR
1523 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1524 inode->i_sb->s_id, inode->i_ino, offset, start);
1530 * Ioend completion routine for merged bios. This can only be called from task
1531 * contexts as merged ioends can be of unbound length. Hence we have to break up
1532 * the writeback completions into manageable chunks to avoid long scheduler
1533 * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1534 * good batch processing throughput without creating adverse scheduler latency
1538 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1540 struct list_head tmp;
1545 list_replace_init(&ioend->io_list, &tmp);
1546 completions = iomap_finish_ioend(ioend, error);
1548 while (!list_empty(&tmp)) {
1549 if (completions > IOEND_BATCH_SIZE * 8) {
1553 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1554 list_del_init(&ioend->io_list);
1555 completions += iomap_finish_ioend(ioend, error);
1558 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1561 * We can merge two adjacent ioends if they have the same set of work to do.
1564 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1566 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1568 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1569 (next->io_flags & IOMAP_F_SHARED))
1571 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1572 (next->io_type == IOMAP_UNWRITTEN))
1574 if (ioend->io_offset + ioend->io_size != next->io_offset)
1577 * Do not merge physically discontiguous ioends. The filesystem
1578 * completion functions will have to iterate the physical
1579 * discontiguities even if we merge the ioends at a logical level, so
1580 * we don't gain anything by merging physical discontiguities here.
1582 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1583 * submission so does not point to the start sector of the bio at
1586 if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1592 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1594 struct iomap_ioend *next;
1596 INIT_LIST_HEAD(&ioend->io_list);
1598 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1600 if (!iomap_ioend_can_merge(ioend, next))
1602 list_move_tail(&next->io_list, &ioend->io_list);
1603 ioend->io_size += next->io_size;
1606 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1609 iomap_ioend_compare(void *priv, const struct list_head *a,
1610 const struct list_head *b)
1612 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1613 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1615 if (ia->io_offset < ib->io_offset)
1617 if (ia->io_offset > ib->io_offset)
1623 iomap_sort_ioends(struct list_head *ioend_list)
1625 list_sort(NULL, ioend_list, iomap_ioend_compare);
1627 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1629 static void iomap_writepage_end_bio(struct bio *bio)
1631 struct iomap_ioend *ioend = bio->bi_private;
1633 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1637 * Submit the final bio for an ioend.
1639 * If @error is non-zero, it means that we have a situation where some part of
1640 * the submission process has failed after we've marked pages for writeback
1641 * and unlocked them. In this situation, we need to fail the bio instead of
1642 * submitting it. This typically only happens on a filesystem shutdown.
1645 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1648 ioend->io_bio->bi_private = ioend;
1649 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1651 if (wpc->ops->prepare_ioend)
1652 error = wpc->ops->prepare_ioend(ioend, error);
1655 * If we're failing the IO now, just mark the ioend with an
1656 * error and finish it. This will run IO completion immediately
1657 * as there is only one reference to the ioend at this point in
1660 ioend->io_bio->bi_status = errno_to_blk_status(error);
1661 bio_endio(ioend->io_bio);
1665 submit_bio(ioend->io_bio);
1669 static struct iomap_ioend *
1670 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1671 loff_t offset, sector_t sector, struct writeback_control *wbc)
1673 struct iomap_ioend *ioend;
1676 bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1677 REQ_OP_WRITE | wbc_to_write_flags(wbc),
1678 GFP_NOFS, &iomap_ioend_bioset);
1679 bio->bi_iter.bi_sector = sector;
1680 wbc_init_bio(wbc, bio);
1682 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1683 INIT_LIST_HEAD(&ioend->io_list);
1684 ioend->io_type = wpc->iomap.type;
1685 ioend->io_flags = wpc->iomap.flags;
1686 ioend->io_inode = inode;
1688 ioend->io_folios = 0;
1689 ioend->io_offset = offset;
1690 ioend->io_bio = bio;
1691 ioend->io_sector = sector;
1696 * Allocate a new bio, and chain the old bio to the new one.
1698 * Note that we have to perform the chaining in this unintuitive order
1699 * so that the bi_private linkage is set up in the right direction for the
1700 * traversal in iomap_finish_ioend().
1703 iomap_chain_bio(struct bio *prev)
1707 new = bio_alloc(prev->bi_bdev, BIO_MAX_VECS, prev->bi_opf, GFP_NOFS);
1708 bio_clone_blkg_association(new, prev);
1709 new->bi_iter.bi_sector = bio_end_sector(prev);
1711 bio_chain(prev, new);
1712 bio_get(prev); /* for iomap_finish_ioend */
1718 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1721 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1722 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1724 if (wpc->iomap.type != wpc->ioend->io_type)
1726 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1728 if (sector != bio_end_sector(wpc->ioend->io_bio))
1731 * Limit ioend bio chain lengths to minimise IO completion latency. This
1732 * also prevents long tight loops ending page writeback on all the
1733 * folios in the ioend.
1735 if (wpc->ioend->io_folios >= IOEND_BATCH_SIZE)
1741 * Test to see if we have an existing ioend structure that we could append to
1742 * first; otherwise finish off the current ioend and start another.
1745 iomap_add_to_ioend(struct inode *inode, loff_t pos, struct folio *folio,
1746 struct iomap_folio_state *ifs, struct iomap_writepage_ctx *wpc,
1747 struct writeback_control *wbc, struct list_head *iolist)
1749 sector_t sector = iomap_sector(&wpc->iomap, pos);
1750 unsigned len = i_blocksize(inode);
1751 size_t poff = offset_in_folio(folio, pos);
1753 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos, sector)) {
1755 list_add(&wpc->ioend->io_list, iolist);
1756 wpc->ioend = iomap_alloc_ioend(inode, wpc, pos, sector, wbc);
1759 if (!bio_add_folio(wpc->ioend->io_bio, folio, len, poff)) {
1760 wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1761 bio_add_folio_nofail(wpc->ioend->io_bio, folio, len, poff);
1765 atomic_add(len, &ifs->write_bytes_pending);
1766 wpc->ioend->io_size += len;
1767 wbc_account_cgroup_owner(wbc, &folio->page, len);
1771 * We implement an immediate ioend submission policy here to avoid needing to
1772 * chain multiple ioends and hence nest mempool allocations which can violate
1773 * the forward progress guarantees we need to provide. The current ioend we're
1774 * adding blocks to is cached in the writepage context, and if the new block
1775 * doesn't append to the cached ioend, it will create a new ioend and cache that
1778 * If a new ioend is created and cached, the old ioend is returned and queued
1779 * locally for submission once the entire page is processed or an error has been
1780 * detected. While ioends are submitted immediately after they are completed,
1781 * batching optimisations are provided by higher level block plugging.
1783 * At the end of a writeback pass, there will be a cached ioend remaining on the
1784 * writepage context that the caller will need to submit.
1787 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1788 struct writeback_control *wbc, struct inode *inode,
1789 struct folio *folio, u64 end_pos)
1791 struct iomap_folio_state *ifs = folio->private;
1792 struct iomap_ioend *ioend, *next;
1793 unsigned len = i_blocksize(inode);
1794 unsigned nblocks = i_blocks_per_folio(inode, folio);
1795 u64 pos = folio_pos(folio);
1796 int error = 0, count = 0, i;
1797 LIST_HEAD(submit_list);
1799 WARN_ON_ONCE(end_pos <= pos);
1801 if (!ifs && nblocks > 1) {
1802 ifs = ifs_alloc(inode, folio, 0);
1803 iomap_set_range_dirty(folio, 0, end_pos - pos);
1806 WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) != 0);
1809 * Walk through the folio to find areas to write back. If we
1810 * run off the end of the current map or find the current map
1811 * invalid, grab a new one.
1813 for (i = 0; i < nblocks && pos < end_pos; i++, pos += len) {
1814 if (ifs && !ifs_block_is_dirty(folio, ifs, i))
1817 error = wpc->ops->map_blocks(wpc, inode, pos);
1820 trace_iomap_writepage_map(inode, &wpc->iomap);
1821 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1823 if (wpc->iomap.type == IOMAP_HOLE)
1825 iomap_add_to_ioend(inode, pos, folio, ifs, wpc, wbc,
1830 wpc->ioend->io_folios++;
1832 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1833 WARN_ON_ONCE(!folio_test_locked(folio));
1834 WARN_ON_ONCE(folio_test_writeback(folio));
1835 WARN_ON_ONCE(folio_test_dirty(folio));
1838 * We cannot cancel the ioend directly here on error. We may have
1839 * already set other pages under writeback and hence we have to run I/O
1840 * completion to mark the error state of the pages under writeback
1843 if (unlikely(error)) {
1845 * Let the filesystem know what portion of the current page
1846 * failed to map. If the page hasn't been added to ioend, it
1847 * won't be affected by I/O completion and we must unlock it
1850 if (wpc->ops->discard_folio)
1851 wpc->ops->discard_folio(folio, pos);
1853 folio_unlock(folio);
1859 * We can have dirty bits set past end of file in page_mkwrite path
1860 * while mapping the last partial folio. Hence it's better to clear
1861 * all the dirty bits in the folio here.
1863 iomap_clear_range_dirty(folio, 0, folio_size(folio));
1864 folio_start_writeback(folio);
1865 folio_unlock(folio);
1868 * Preserve the original error if there was one; catch
1869 * submission errors here and propagate into subsequent ioend
1872 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1875 list_del_init(&ioend->io_list);
1876 error2 = iomap_submit_ioend(wpc, ioend, error);
1877 if (error2 && !error)
1882 * We can end up here with no error and nothing to write only if we race
1883 * with a partial page truncate on a sub-page block sized filesystem.
1886 folio_end_writeback(folio);
1888 mapping_set_error(inode->i_mapping, error);
1893 * Write out a dirty page.
1895 * For delalloc space on the page, we need to allocate space and flush it.
1896 * For unwritten space on the page, we need to start the conversion to
1897 * regular allocated space.
1899 static int iomap_do_writepage(struct folio *folio,
1900 struct writeback_control *wbc, void *data)
1902 struct iomap_writepage_ctx *wpc = data;
1903 struct inode *inode = folio->mapping->host;
1906 trace_iomap_writepage(inode, folio_pos(folio), folio_size(folio));
1909 * Refuse to write the folio out if we're called from reclaim context.
1911 * This avoids stack overflows when called from deeply used stacks in
1912 * random callers for direct reclaim or memcg reclaim. We explicitly
1913 * allow reclaim from kswapd as the stack usage there is relatively low.
1915 * This should never happen except in the case of a VM regression so
1918 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1923 * Is this folio beyond the end of the file?
1925 * The folio index is less than the end_index, adjust the end_pos
1926 * to the highest offset that this folio should represent.
1927 * -----------------------------------------------------
1928 * | file mapping | <EOF> |
1929 * -----------------------------------------------------
1930 * | Page ... | Page N-2 | Page N-1 | Page N | |
1931 * ^--------------------------------^----------|--------
1932 * | desired writeback range | see else |
1933 * ---------------------------------^------------------|
1935 isize = i_size_read(inode);
1936 end_pos = folio_pos(folio) + folio_size(folio);
1937 if (end_pos > isize) {
1939 * Check whether the page to write out is beyond or straddles
1941 * -------------------------------------------------------
1942 * | file mapping | <EOF> |
1943 * -------------------------------------------------------
1944 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1945 * ^--------------------------------^-----------|---------
1947 * ---------------------------------^-----------|--------|
1949 size_t poff = offset_in_folio(folio, isize);
1950 pgoff_t end_index = isize >> PAGE_SHIFT;
1953 * Skip the page if it's fully outside i_size, e.g.
1954 * due to a truncate operation that's in progress. We've
1955 * cleaned this page and truncate will finish things off for
1958 * Note that the end_index is unsigned long. If the given
1959 * offset is greater than 16TB on a 32-bit system then if we
1960 * checked if the page is fully outside i_size with
1961 * "if (page->index >= end_index + 1)", "end_index + 1" would
1962 * overflow and evaluate to 0. Hence this page would be
1963 * redirtied and written out repeatedly, which would result in
1964 * an infinite loop; the user program performing this operation
1965 * would hang. Instead, we can detect this situation by
1966 * checking if the page is totally beyond i_size or if its
1967 * offset is just equal to the EOF.
1969 if (folio->index > end_index ||
1970 (folio->index == end_index && poff == 0))
1974 * The page straddles i_size. It must be zeroed out on each
1975 * and every writepage invocation because it may be mmapped.
1976 * "A file is mapped in multiples of the page size. For a file
1977 * that is not a multiple of the page size, the remaining
1978 * memory is zeroed when mapped, and writes to that region are
1979 * not written out to the file."
1981 folio_zero_segment(folio, poff, folio_size(folio));
1985 return iomap_writepage_map(wpc, wbc, inode, folio, end_pos);
1988 folio_redirty_for_writepage(wbc, folio);
1990 folio_unlock(folio);
1995 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1996 struct iomap_writepage_ctx *wpc,
1997 const struct iomap_writeback_ops *ops)
2002 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
2005 return iomap_submit_ioend(wpc, wpc->ioend, ret);
2007 EXPORT_SYMBOL_GPL(iomap_writepages);
2009 static int __init iomap_init(void)
2011 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
2012 offsetof(struct iomap_ioend, io_inline_bio),
2015 fs_initcall(iomap_init);