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
27 * Structure allocated for each folio when block size < folio size
28 * to track sub-folio uptodate status and I/O completions.
31 atomic_t read_bytes_pending;
32 atomic_t write_bytes_pending;
33 spinlock_t uptodate_lock;
34 unsigned long uptodate[];
37 static inline struct iomap_page *to_iomap_page(struct folio *folio)
39 if (folio_test_private(folio))
40 return folio_get_private(folio);
44 static struct bio_set iomap_ioend_bioset;
46 static struct iomap_page *
47 iomap_page_create(struct inode *inode, struct folio *folio)
49 struct iomap_page *iop = to_iomap_page(folio);
50 unsigned int nr_blocks = i_blocks_per_folio(inode, folio);
52 if (iop || nr_blocks <= 1)
55 iop = kzalloc(struct_size(iop, uptodate, BITS_TO_LONGS(nr_blocks)),
56 GFP_NOFS | __GFP_NOFAIL);
57 spin_lock_init(&iop->uptodate_lock);
58 if (folio_test_uptodate(folio))
59 bitmap_fill(iop->uptodate, nr_blocks);
60 folio_attach_private(folio, iop);
64 static void iomap_page_release(struct folio *folio)
66 struct iomap_page *iop = folio_detach_private(folio);
67 struct inode *inode = folio->mapping->host;
68 unsigned int nr_blocks = i_blocks_per_folio(inode, folio);
72 WARN_ON_ONCE(atomic_read(&iop->read_bytes_pending));
73 WARN_ON_ONCE(atomic_read(&iop->write_bytes_pending));
74 WARN_ON_ONCE(bitmap_full(iop->uptodate, nr_blocks) !=
75 folio_test_uptodate(folio));
80 * Calculate the range inside the folio that we actually need to read.
82 static void iomap_adjust_read_range(struct inode *inode, struct folio *folio,
83 loff_t *pos, loff_t length, size_t *offp, size_t *lenp)
85 struct iomap_page *iop = to_iomap_page(folio);
86 loff_t orig_pos = *pos;
87 loff_t isize = i_size_read(inode);
88 unsigned block_bits = inode->i_blkbits;
89 unsigned block_size = (1 << block_bits);
90 size_t poff = offset_in_folio(folio, *pos);
91 size_t plen = min_t(loff_t, folio_size(folio) - poff, length);
92 unsigned first = poff >> block_bits;
93 unsigned last = (poff + plen - 1) >> block_bits;
96 * If the block size is smaller than the page size, we need to check the
97 * per-block uptodate status and adjust the offset and length if needed
98 * to avoid reading in already uptodate ranges.
103 /* move forward for each leading block marked uptodate */
104 for (i = first; i <= last; i++) {
105 if (!test_bit(i, iop->uptodate))
113 /* truncate len if we find any trailing uptodate block(s) */
114 for ( ; i <= last; i++) {
115 if (test_bit(i, iop->uptodate)) {
116 plen -= (last - i + 1) * block_size;
124 * If the extent spans the block that contains the i_size, we need to
125 * handle both halves separately so that we properly zero data in the
126 * page cache for blocks that are entirely outside of i_size.
128 if (orig_pos <= isize && orig_pos + length > isize) {
129 unsigned end = offset_in_folio(folio, isize - 1) >> block_bits;
131 if (first <= end && last > end)
132 plen -= (last - end) * block_size;
139 static void iomap_iop_set_range_uptodate(struct folio *folio,
140 struct iomap_page *iop, size_t off, size_t len)
142 struct inode *inode = folio->mapping->host;
143 unsigned first = off >> inode->i_blkbits;
144 unsigned last = (off + len - 1) >> inode->i_blkbits;
147 spin_lock_irqsave(&iop->uptodate_lock, flags);
148 bitmap_set(iop->uptodate, first, last - first + 1);
149 if (bitmap_full(iop->uptodate, i_blocks_per_folio(inode, folio)))
150 folio_mark_uptodate(folio);
151 spin_unlock_irqrestore(&iop->uptodate_lock, flags);
154 static void iomap_set_range_uptodate(struct folio *folio,
155 struct iomap_page *iop, size_t off, size_t len)
157 if (folio_test_error(folio))
161 iomap_iop_set_range_uptodate(folio, iop, off, len);
163 folio_mark_uptodate(folio);
166 static void iomap_finish_folio_read(struct folio *folio, size_t offset,
167 size_t len, int error)
169 struct iomap_page *iop = to_iomap_page(folio);
171 if (unlikely(error)) {
172 folio_clear_uptodate(folio);
173 folio_set_error(folio);
175 iomap_set_range_uptodate(folio, iop, offset, len);
178 if (!iop || atomic_sub_and_test(len, &iop->read_bytes_pending))
182 static void iomap_read_end_io(struct bio *bio)
184 int error = blk_status_to_errno(bio->bi_status);
185 struct folio_iter fi;
187 bio_for_each_folio_all(fi, bio)
188 iomap_finish_folio_read(fi.folio, fi.offset, fi.length, error);
192 struct iomap_readpage_ctx {
193 struct folio *cur_folio;
194 bool cur_folio_in_bio;
196 struct readahead_control *rac;
200 * iomap_read_inline_data - copy inline data into the page cache
201 * @iter: iteration structure
202 * @folio: folio to copy to
204 * Copy the inline data in @iter into @folio and zero out the rest of the folio.
205 * Only a single IOMAP_INLINE extent is allowed at the end of each file.
206 * Returns zero for success to complete the read, or the usual negative errno.
208 static int iomap_read_inline_data(const struct iomap_iter *iter,
211 struct iomap_page *iop;
212 const struct iomap *iomap = iomap_iter_srcmap(iter);
213 size_t size = i_size_read(iter->inode) - iomap->offset;
214 size_t poff = offset_in_page(iomap->offset);
215 size_t offset = offset_in_folio(folio, iomap->offset);
218 if (folio_test_uptodate(folio))
221 if (WARN_ON_ONCE(size > PAGE_SIZE - poff))
223 if (WARN_ON_ONCE(size > PAGE_SIZE -
224 offset_in_page(iomap->inline_data)))
226 if (WARN_ON_ONCE(size > iomap->length))
229 iop = iomap_page_create(iter->inode, folio);
231 iop = to_iomap_page(folio);
233 addr = kmap_local_folio(folio, offset);
234 memcpy(addr, iomap->inline_data, size);
235 memset(addr + size, 0, PAGE_SIZE - poff - size);
237 iomap_set_range_uptodate(folio, iop, offset, PAGE_SIZE - poff);
241 static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
244 const struct iomap *srcmap = iomap_iter_srcmap(iter);
246 return srcmap->type != IOMAP_MAPPED ||
247 (srcmap->flags & IOMAP_F_NEW) ||
248 pos >= i_size_read(iter->inode);
251 static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
252 struct iomap_readpage_ctx *ctx, loff_t offset)
254 const struct iomap *iomap = &iter->iomap;
255 loff_t pos = iter->pos + offset;
256 loff_t length = iomap_length(iter) - offset;
257 struct folio *folio = ctx->cur_folio;
258 struct iomap_page *iop;
259 loff_t orig_pos = pos;
263 if (iomap->type == IOMAP_INLINE)
264 return iomap_read_inline_data(iter, folio);
266 /* zero post-eof blocks as the page may be mapped */
267 iop = iomap_page_create(iter->inode, folio);
268 iomap_adjust_read_range(iter->inode, folio, &pos, length, &poff, &plen);
272 if (iomap_block_needs_zeroing(iter, pos)) {
273 folio_zero_range(folio, poff, plen);
274 iomap_set_range_uptodate(folio, iop, poff, plen);
278 ctx->cur_folio_in_bio = true;
280 atomic_add(plen, &iop->read_bytes_pending);
282 sector = iomap_sector(iomap, pos);
284 bio_end_sector(ctx->bio) != sector ||
285 !bio_add_folio(ctx->bio, folio, plen, poff)) {
286 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
287 gfp_t orig_gfp = gfp;
288 unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
291 submit_bio(ctx->bio);
293 if (ctx->rac) /* same as readahead_gfp_mask */
294 gfp |= __GFP_NORETRY | __GFP_NOWARN;
295 ctx->bio = bio_alloc(iomap->bdev, bio_max_segs(nr_vecs),
298 * If the bio_alloc fails, try it again for a single page to
299 * avoid having to deal with partial page reads. This emulates
300 * what do_mpage_readpage does.
303 ctx->bio = bio_alloc(iomap->bdev, 1, REQ_OP_READ,
307 ctx->bio->bi_opf |= REQ_RAHEAD;
308 ctx->bio->bi_iter.bi_sector = sector;
309 ctx->bio->bi_end_io = iomap_read_end_io;
310 bio_add_folio(ctx->bio, folio, plen, poff);
315 * Move the caller beyond our range so that it keeps making progress.
316 * For that, we have to include any leading non-uptodate ranges, but
317 * we can skip trailing ones as they will be handled in the next
320 return pos - orig_pos + plen;
324 iomap_readpage(struct page *page, const struct iomap_ops *ops)
326 struct folio *folio = page_folio(page);
327 struct iomap_iter iter = {
328 .inode = folio->mapping->host,
329 .pos = folio_pos(folio),
330 .len = folio_size(folio),
332 struct iomap_readpage_ctx ctx = {
337 trace_iomap_readpage(iter.inode, 1);
339 while ((ret = iomap_iter(&iter, ops)) > 0)
340 iter.processed = iomap_readpage_iter(&iter, &ctx, 0);
343 folio_set_error(folio);
347 WARN_ON_ONCE(!ctx.cur_folio_in_bio);
349 WARN_ON_ONCE(ctx.cur_folio_in_bio);
354 * Just like mpage_readahead and block_read_full_page, we always
355 * return 0 and just mark the page as PageError on errors. This
356 * should be cleaned up throughout the stack eventually.
360 EXPORT_SYMBOL_GPL(iomap_readpage);
362 static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
363 struct iomap_readpage_ctx *ctx)
365 loff_t length = iomap_length(iter);
368 for (done = 0; done < length; done += ret) {
369 if (ctx->cur_folio &&
370 offset_in_folio(ctx->cur_folio, iter->pos + done) == 0) {
371 if (!ctx->cur_folio_in_bio)
372 folio_unlock(ctx->cur_folio);
373 ctx->cur_folio = NULL;
375 if (!ctx->cur_folio) {
376 ctx->cur_folio = readahead_folio(ctx->rac);
377 ctx->cur_folio_in_bio = false;
379 ret = iomap_readpage_iter(iter, ctx, done);
388 * iomap_readahead - Attempt to read pages from a file.
389 * @rac: Describes the pages to be read.
390 * @ops: The operations vector for the filesystem.
392 * This function is for filesystems to call to implement their readahead
393 * address_space operation.
395 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
396 * blocks from disc), and may wait for it. The caller may be trying to
397 * access a different page, and so sleeping excessively should be avoided.
398 * It may allocate memory, but should avoid costly allocations. This
399 * function is called with memalloc_nofs set, so allocations will not cause
400 * the filesystem to be reentered.
402 void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
404 struct iomap_iter iter = {
405 .inode = rac->mapping->host,
406 .pos = readahead_pos(rac),
407 .len = readahead_length(rac),
409 struct iomap_readpage_ctx ctx = {
413 trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
415 while (iomap_iter(&iter, ops) > 0)
416 iter.processed = iomap_readahead_iter(&iter, &ctx);
421 if (!ctx.cur_folio_in_bio)
422 folio_unlock(ctx.cur_folio);
425 EXPORT_SYMBOL_GPL(iomap_readahead);
428 * iomap_is_partially_uptodate checks whether blocks within a folio are
431 * Returns true if all blocks which correspond to the specified part
432 * of the folio are uptodate.
434 bool iomap_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
436 struct iomap_page *iop = to_iomap_page(folio);
437 struct inode *inode = folio->mapping->host;
438 unsigned first, last, i;
443 /* Caller's range may extend past the end of this folio */
444 count = min(folio_size(folio) - from, count);
446 /* First and last blocks in range within folio */
447 first = from >> inode->i_blkbits;
448 last = (from + count - 1) >> inode->i_blkbits;
450 for (i = first; i <= last; i++)
451 if (!test_bit(i, iop->uptodate))
455 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
458 iomap_releasepage(struct page *page, gfp_t gfp_mask)
460 struct folio *folio = page_folio(page);
462 trace_iomap_releasepage(folio->mapping->host, folio_pos(folio),
466 * mm accommodates an old ext3 case where clean pages might not have had
467 * the dirty bit cleared. Thus, it can send actual dirty pages to
468 * ->releasepage() via shrink_active_list(); skip those here.
470 if (folio_test_dirty(folio) || folio_test_writeback(folio))
472 iomap_page_release(folio);
475 EXPORT_SYMBOL_GPL(iomap_releasepage);
477 void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
479 trace_iomap_invalidate_folio(folio->mapping->host,
480 folio_pos(folio) + offset, len);
483 * If we're invalidating the entire folio, clear the dirty state
484 * from it and release it to avoid unnecessary buildup of the LRU.
486 if (offset == 0 && len == folio_size(folio)) {
487 WARN_ON_ONCE(folio_test_writeback(folio));
488 folio_cancel_dirty(folio);
489 iomap_page_release(folio);
490 } else if (folio_test_large(folio)) {
491 /* Must release the iop so the page can be split */
492 WARN_ON_ONCE(!folio_test_uptodate(folio) &&
493 folio_test_dirty(folio));
494 iomap_page_release(folio);
497 EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
499 #ifdef CONFIG_MIGRATION
501 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
502 struct page *page, enum migrate_mode mode)
504 struct folio *folio = page_folio(page);
505 struct folio *newfolio = page_folio(newpage);
508 ret = folio_migrate_mapping(mapping, newfolio, folio, 0);
509 if (ret != MIGRATEPAGE_SUCCESS)
512 if (folio_test_private(folio))
513 folio_attach_private(newfolio, folio_detach_private(folio));
515 if (mode != MIGRATE_SYNC_NO_COPY)
516 folio_migrate_copy(newfolio, folio);
518 folio_migrate_flags(newfolio, folio);
519 return MIGRATEPAGE_SUCCESS;
521 EXPORT_SYMBOL_GPL(iomap_migrate_page);
522 #endif /* CONFIG_MIGRATION */
525 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
527 loff_t i_size = i_size_read(inode);
530 * Only truncate newly allocated pages beyoned EOF, even if the
531 * write started inside the existing inode size.
533 if (pos + len > i_size)
534 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
537 static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
538 size_t poff, size_t plen, const struct iomap *iomap)
543 bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
544 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
545 bio_add_folio(&bio, folio, plen, poff);
546 return submit_bio_wait(&bio);
549 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
550 size_t len, struct folio *folio)
552 const struct iomap *srcmap = iomap_iter_srcmap(iter);
553 struct iomap_page *iop = iomap_page_create(iter->inode, folio);
554 loff_t block_size = i_blocksize(iter->inode);
555 loff_t block_start = round_down(pos, block_size);
556 loff_t block_end = round_up(pos + len, block_size);
557 size_t from = offset_in_folio(folio, pos), to = from + len;
560 if (folio_test_uptodate(folio))
562 folio_clear_error(folio);
565 iomap_adjust_read_range(iter->inode, folio, &block_start,
566 block_end - block_start, &poff, &plen);
570 if (!(iter->flags & IOMAP_UNSHARE) &&
571 (from <= poff || from >= poff + plen) &&
572 (to <= poff || to >= poff + plen))
575 if (iomap_block_needs_zeroing(iter, block_start)) {
576 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
578 folio_zero_segments(folio, poff, from, to, poff + plen);
580 int status = iomap_read_folio_sync(block_start, folio,
585 iomap_set_range_uptodate(folio, iop, poff, plen);
586 } while ((block_start += plen) < block_end);
591 static int iomap_write_begin_inline(const struct iomap_iter *iter,
594 /* needs more work for the tailpacking case; disable for now */
595 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
597 return iomap_read_inline_data(iter, folio);
600 static int iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
601 size_t len, struct folio **foliop)
603 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
604 const struct iomap *srcmap = iomap_iter_srcmap(iter);
606 unsigned fgp = FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE | FGP_NOFS;
609 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
610 if (srcmap != &iter->iomap)
611 BUG_ON(pos + len > srcmap->offset + srcmap->length);
613 if (fatal_signal_pending(current))
616 if (!mapping_large_folio_support(iter->inode->i_mapping))
617 len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
619 if (page_ops && page_ops->page_prepare) {
620 status = page_ops->page_prepare(iter->inode, pos, len);
625 folio = __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
626 fgp, mapping_gfp_mask(iter->inode->i_mapping));
631 if (pos + len > folio_pos(folio) + folio_size(folio))
632 len = folio_pos(folio) + folio_size(folio) - pos;
634 if (srcmap->type == IOMAP_INLINE)
635 status = iomap_write_begin_inline(iter, folio);
636 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
637 status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
639 status = __iomap_write_begin(iter, pos, len, folio);
641 if (unlikely(status))
650 iomap_write_failed(iter->inode, pos, len);
653 if (page_ops && page_ops->page_done)
654 page_ops->page_done(iter->inode, pos, 0, NULL);
658 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
659 size_t copied, struct folio *folio)
661 struct iomap_page *iop = to_iomap_page(folio);
662 flush_dcache_folio(folio);
665 * The blocks that were entirely written will now be uptodate, so we
666 * don't have to worry about a readpage reading them and overwriting a
667 * partial write. However, if we've encountered a short write and only
668 * partially written into a block, it will not be marked uptodate, so a
669 * readpage might come in and destroy our partial write.
671 * Do the simplest thing and just treat any short write to a
672 * non-uptodate page as a zero-length write, and force the caller to
673 * redo the whole thing.
675 if (unlikely(copied < len && !folio_test_uptodate(folio)))
677 iomap_set_range_uptodate(folio, iop, offset_in_folio(folio, pos), len);
678 filemap_dirty_folio(inode->i_mapping, folio);
682 static size_t iomap_write_end_inline(const struct iomap_iter *iter,
683 struct folio *folio, loff_t pos, size_t copied)
685 const struct iomap *iomap = &iter->iomap;
688 WARN_ON_ONCE(!folio_test_uptodate(folio));
689 BUG_ON(!iomap_inline_data_valid(iomap));
691 flush_dcache_folio(folio);
692 addr = kmap_local_folio(folio, pos);
693 memcpy(iomap_inline_data(iomap, pos), addr, copied);
696 mark_inode_dirty(iter->inode);
700 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
701 static size_t iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
702 size_t copied, struct folio *folio)
704 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
705 const struct iomap *srcmap = iomap_iter_srcmap(iter);
706 loff_t old_size = iter->inode->i_size;
709 if (srcmap->type == IOMAP_INLINE) {
710 ret = iomap_write_end_inline(iter, folio, pos, copied);
711 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
712 ret = block_write_end(NULL, iter->inode->i_mapping, pos, len,
713 copied, &folio->page, NULL);
715 ret = __iomap_write_end(iter->inode, pos, len, copied, folio);
719 * Update the in-memory inode size after copying the data into the page
720 * cache. It's up to the file system to write the updated size to disk,
721 * preferably after I/O completion so that no stale data is exposed.
723 if (pos + ret > old_size) {
724 i_size_write(iter->inode, pos + ret);
725 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
730 pagecache_isize_extended(iter->inode, old_size, pos);
731 if (page_ops && page_ops->page_done)
732 page_ops->page_done(iter->inode, pos, ret, &folio->page);
736 iomap_write_failed(iter->inode, pos, len);
740 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
742 loff_t length = iomap_length(iter);
743 loff_t pos = iter->pos;
750 unsigned long offset; /* Offset into pagecache page */
751 unsigned long bytes; /* Bytes to write to page */
752 size_t copied; /* Bytes copied from user */
754 offset = offset_in_page(pos);
755 bytes = min_t(unsigned long, PAGE_SIZE - offset,
762 * Bring in the user page that we'll copy from _first_.
763 * Otherwise there's a nasty deadlock on copying from the
764 * same page as we're writing to, without it being marked
767 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
772 status = iomap_write_begin(iter, pos, bytes, &folio);
773 if (unlikely(status))
776 page = folio_file_page(folio, pos >> PAGE_SHIFT);
777 if (mapping_writably_mapped(iter->inode->i_mapping))
778 flush_dcache_page(page);
780 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
782 status = iomap_write_end(iter, pos, bytes, copied, folio);
784 if (unlikely(copied != status))
785 iov_iter_revert(i, copied - status);
788 if (unlikely(status == 0)) {
790 * A short copy made iomap_write_end() reject the
791 * thing entirely. Might be memory poisoning
792 * halfway through, might be a race with munmap,
793 * might be severe memory pressure.
803 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
804 } while (iov_iter_count(i) && length);
806 return written ? written : status;
810 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
811 const struct iomap_ops *ops)
813 struct iomap_iter iter = {
814 .inode = iocb->ki_filp->f_mapping->host,
816 .len = iov_iter_count(i),
817 .flags = IOMAP_WRITE,
821 while ((ret = iomap_iter(&iter, ops)) > 0)
822 iter.processed = iomap_write_iter(&iter, i);
823 if (iter.pos == iocb->ki_pos)
825 return iter.pos - iocb->ki_pos;
827 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
829 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
831 struct iomap *iomap = &iter->iomap;
832 const struct iomap *srcmap = iomap_iter_srcmap(iter);
833 loff_t pos = iter->pos;
834 loff_t length = iomap_length(iter);
838 /* don't bother with blocks that are not shared to start with */
839 if (!(iomap->flags & IOMAP_F_SHARED))
841 /* don't bother with holes or unwritten extents */
842 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
846 unsigned long offset = offset_in_page(pos);
847 unsigned long bytes = min_t(loff_t, PAGE_SIZE - offset, length);
850 status = iomap_write_begin(iter, pos, bytes, &folio);
851 if (unlikely(status))
854 status = iomap_write_end(iter, pos, bytes, bytes, folio);
855 if (WARN_ON_ONCE(status == 0))
864 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
871 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
872 const struct iomap_ops *ops)
874 struct iomap_iter iter = {
878 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
882 while ((ret = iomap_iter(&iter, ops)) > 0)
883 iter.processed = iomap_unshare_iter(&iter);
886 EXPORT_SYMBOL_GPL(iomap_file_unshare);
888 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
890 const struct iomap *srcmap = iomap_iter_srcmap(iter);
891 loff_t pos = iter->pos;
892 loff_t length = iomap_length(iter);
895 /* already zeroed? we're done. */
896 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
903 size_t bytes = min_t(u64, SIZE_MAX, length);
905 status = iomap_write_begin(iter, pos, bytes, &folio);
909 offset = offset_in_folio(folio, pos);
910 if (bytes > folio_size(folio) - offset)
911 bytes = folio_size(folio) - offset;
913 folio_zero_range(folio, offset, bytes);
914 folio_mark_accessed(folio);
916 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
917 if (WARN_ON_ONCE(bytes == 0))
925 } while (length > 0);
931 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
932 const struct iomap_ops *ops)
934 struct iomap_iter iter = {
942 while ((ret = iomap_iter(&iter, ops)) > 0)
943 iter.processed = iomap_zero_iter(&iter, did_zero);
946 EXPORT_SYMBOL_GPL(iomap_zero_range);
949 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
950 const struct iomap_ops *ops)
952 unsigned int blocksize = i_blocksize(inode);
953 unsigned int off = pos & (blocksize - 1);
955 /* Block boundary? Nothing to do */
958 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
960 EXPORT_SYMBOL_GPL(iomap_truncate_page);
962 static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
965 loff_t length = iomap_length(iter);
968 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
969 ret = __block_write_begin_int(folio, iter->pos, length, NULL,
973 block_commit_write(&folio->page, 0, length);
975 WARN_ON_ONCE(!folio_test_uptodate(folio));
976 folio_mark_dirty(folio);
982 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
984 struct iomap_iter iter = {
985 .inode = file_inode(vmf->vma->vm_file),
986 .flags = IOMAP_WRITE | IOMAP_FAULT,
988 struct folio *folio = page_folio(vmf->page);
992 ret = folio_mkwrite_check_truncate(folio, iter.inode);
995 iter.pos = folio_pos(folio);
997 while ((ret = iomap_iter(&iter, ops)) > 0)
998 iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1002 folio_wait_stable(folio);
1003 return VM_FAULT_LOCKED;
1005 folio_unlock(folio);
1006 return block_page_mkwrite_return(ret);
1008 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1010 static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1011 size_t len, int error)
1013 struct iomap_page *iop = to_iomap_page(folio);
1016 folio_set_error(folio);
1017 mapping_set_error(inode->i_mapping, error);
1020 WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !iop);
1021 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) <= 0);
1023 if (!iop || atomic_sub_and_test(len, &iop->write_bytes_pending))
1024 folio_end_writeback(folio);
1028 * We're now finished for good with this ioend structure. Update the page
1029 * state, release holds on bios, and finally free up memory. Do not use the
1033 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1035 struct inode *inode = ioend->io_inode;
1036 struct bio *bio = &ioend->io_inline_bio;
1037 struct bio *last = ioend->io_bio, *next;
1038 u64 start = bio->bi_iter.bi_sector;
1039 loff_t offset = ioend->io_offset;
1040 bool quiet = bio_flagged(bio, BIO_QUIET);
1041 u32 folio_count = 0;
1043 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1044 struct folio_iter fi;
1047 * For the last bio, bi_private points to the ioend, so we
1048 * need to explicitly end the iteration here.
1053 next = bio->bi_private;
1055 /* walk all folios in bio, ending page IO on them */
1056 bio_for_each_folio_all(fi, bio) {
1057 iomap_finish_folio_write(inode, fi.folio, fi.length,
1063 /* The ioend has been freed by bio_put() */
1065 if (unlikely(error && !quiet)) {
1066 printk_ratelimited(KERN_ERR
1067 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1068 inode->i_sb->s_id, inode->i_ino, offset, start);
1074 * Ioend completion routine for merged bios. This can only be called from task
1075 * contexts as merged ioends can be of unbound length. Hence we have to break up
1076 * the writeback completions into manageable chunks to avoid long scheduler
1077 * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1078 * good batch processing throughput without creating adverse scheduler latency
1082 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1084 struct list_head tmp;
1089 list_replace_init(&ioend->io_list, &tmp);
1090 completions = iomap_finish_ioend(ioend, error);
1092 while (!list_empty(&tmp)) {
1093 if (completions > IOEND_BATCH_SIZE * 8) {
1097 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1098 list_del_init(&ioend->io_list);
1099 completions += iomap_finish_ioend(ioend, error);
1102 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1105 * We can merge two adjacent ioends if they have the same set of work to do.
1108 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1110 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1112 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1113 (next->io_flags & IOMAP_F_SHARED))
1115 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1116 (next->io_type == IOMAP_UNWRITTEN))
1118 if (ioend->io_offset + ioend->io_size != next->io_offset)
1121 * Do not merge physically discontiguous ioends. The filesystem
1122 * completion functions will have to iterate the physical
1123 * discontiguities even if we merge the ioends at a logical level, so
1124 * we don't gain anything by merging physical discontiguities here.
1126 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1127 * submission so does not point to the start sector of the bio at
1130 if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1136 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1138 struct iomap_ioend *next;
1140 INIT_LIST_HEAD(&ioend->io_list);
1142 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1144 if (!iomap_ioend_can_merge(ioend, next))
1146 list_move_tail(&next->io_list, &ioend->io_list);
1147 ioend->io_size += next->io_size;
1150 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1153 iomap_ioend_compare(void *priv, const struct list_head *a,
1154 const struct list_head *b)
1156 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1157 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1159 if (ia->io_offset < ib->io_offset)
1161 if (ia->io_offset > ib->io_offset)
1167 iomap_sort_ioends(struct list_head *ioend_list)
1169 list_sort(NULL, ioend_list, iomap_ioend_compare);
1171 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1173 static void iomap_writepage_end_bio(struct bio *bio)
1175 struct iomap_ioend *ioend = bio->bi_private;
1177 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1181 * Submit the final bio for an ioend.
1183 * If @error is non-zero, it means that we have a situation where some part of
1184 * the submission process has failed after we've marked pages for writeback
1185 * and unlocked them. In this situation, we need to fail the bio instead of
1186 * submitting it. This typically only happens on a filesystem shutdown.
1189 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1192 ioend->io_bio->bi_private = ioend;
1193 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1195 if (wpc->ops->prepare_ioend)
1196 error = wpc->ops->prepare_ioend(ioend, error);
1199 * If we're failing the IO now, just mark the ioend with an
1200 * error and finish it. This will run IO completion immediately
1201 * as there is only one reference to the ioend at this point in
1204 ioend->io_bio->bi_status = errno_to_blk_status(error);
1205 bio_endio(ioend->io_bio);
1209 submit_bio(ioend->io_bio);
1213 static struct iomap_ioend *
1214 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1215 loff_t offset, sector_t sector, struct writeback_control *wbc)
1217 struct iomap_ioend *ioend;
1220 bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1221 REQ_OP_WRITE | wbc_to_write_flags(wbc),
1222 GFP_NOFS, &iomap_ioend_bioset);
1223 bio->bi_iter.bi_sector = sector;
1224 wbc_init_bio(wbc, bio);
1226 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1227 INIT_LIST_HEAD(&ioend->io_list);
1228 ioend->io_type = wpc->iomap.type;
1229 ioend->io_flags = wpc->iomap.flags;
1230 ioend->io_inode = inode;
1232 ioend->io_folios = 0;
1233 ioend->io_offset = offset;
1234 ioend->io_bio = bio;
1235 ioend->io_sector = sector;
1240 * Allocate a new bio, and chain the old bio to the new one.
1242 * Note that we have to perform the chaining in this unintuitive order
1243 * so that the bi_private linkage is set up in the right direction for the
1244 * traversal in iomap_finish_ioend().
1247 iomap_chain_bio(struct bio *prev)
1251 new = bio_alloc(prev->bi_bdev, BIO_MAX_VECS, prev->bi_opf, GFP_NOFS);
1252 bio_clone_blkg_association(new, prev);
1253 new->bi_iter.bi_sector = bio_end_sector(prev);
1255 bio_chain(prev, new);
1256 bio_get(prev); /* for iomap_finish_ioend */
1262 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1265 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1266 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1268 if (wpc->iomap.type != wpc->ioend->io_type)
1270 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1272 if (sector != bio_end_sector(wpc->ioend->io_bio))
1275 * Limit ioend bio chain lengths to minimise IO completion latency. This
1276 * also prevents long tight loops ending page writeback on all the
1277 * folios in the ioend.
1279 if (wpc->ioend->io_folios >= IOEND_BATCH_SIZE)
1285 * Test to see if we have an existing ioend structure that we could append to
1286 * first; otherwise finish off the current ioend and start another.
1289 iomap_add_to_ioend(struct inode *inode, loff_t pos, struct folio *folio,
1290 struct iomap_page *iop, struct iomap_writepage_ctx *wpc,
1291 struct writeback_control *wbc, struct list_head *iolist)
1293 sector_t sector = iomap_sector(&wpc->iomap, pos);
1294 unsigned len = i_blocksize(inode);
1295 size_t poff = offset_in_folio(folio, pos);
1297 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos, sector)) {
1299 list_add(&wpc->ioend->io_list, iolist);
1300 wpc->ioend = iomap_alloc_ioend(inode, wpc, pos, sector, wbc);
1303 if (!bio_add_folio(wpc->ioend->io_bio, folio, len, poff)) {
1304 wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1305 bio_add_folio(wpc->ioend->io_bio, folio, len, poff);
1309 atomic_add(len, &iop->write_bytes_pending);
1310 wpc->ioend->io_size += len;
1311 wbc_account_cgroup_owner(wbc, &folio->page, len);
1315 * We implement an immediate ioend submission policy here to avoid needing to
1316 * chain multiple ioends and hence nest mempool allocations which can violate
1317 * the forward progress guarantees we need to provide. The current ioend we're
1318 * adding blocks to is cached in the writepage context, and if the new block
1319 * doesn't append to the cached ioend, it will create a new ioend and cache that
1322 * If a new ioend is created and cached, the old ioend is returned and queued
1323 * locally for submission once the entire page is processed or an error has been
1324 * detected. While ioends are submitted immediately after they are completed,
1325 * batching optimisations are provided by higher level block plugging.
1327 * At the end of a writeback pass, there will be a cached ioend remaining on the
1328 * writepage context that the caller will need to submit.
1331 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1332 struct writeback_control *wbc, struct inode *inode,
1333 struct folio *folio, u64 end_pos)
1335 struct iomap_page *iop = iomap_page_create(inode, folio);
1336 struct iomap_ioend *ioend, *next;
1337 unsigned len = i_blocksize(inode);
1338 unsigned nblocks = i_blocks_per_folio(inode, folio);
1339 u64 pos = folio_pos(folio);
1340 int error = 0, count = 0, i;
1341 LIST_HEAD(submit_list);
1343 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) != 0);
1346 * Walk through the folio to find areas to write back. If we
1347 * run off the end of the current map or find the current map
1348 * invalid, grab a new one.
1350 for (i = 0; i < nblocks && pos < end_pos; i++, pos += len) {
1351 if (iop && !test_bit(i, iop->uptodate))
1354 error = wpc->ops->map_blocks(wpc, inode, pos);
1357 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1359 if (wpc->iomap.type == IOMAP_HOLE)
1361 iomap_add_to_ioend(inode, pos, folio, iop, wpc, wbc,
1366 wpc->ioend->io_folios++;
1368 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1369 WARN_ON_ONCE(!folio_test_locked(folio));
1370 WARN_ON_ONCE(folio_test_writeback(folio));
1371 WARN_ON_ONCE(folio_test_dirty(folio));
1374 * We cannot cancel the ioend directly here on error. We may have
1375 * already set other pages under writeback and hence we have to run I/O
1376 * completion to mark the error state of the pages under writeback
1379 if (unlikely(error)) {
1381 * Let the filesystem know what portion of the current page
1382 * failed to map. If the page hasn't been added to ioend, it
1383 * won't be affected by I/O completion and we must unlock it
1386 if (wpc->ops->discard_folio)
1387 wpc->ops->discard_folio(folio, pos);
1389 folio_clear_uptodate(folio);
1390 folio_unlock(folio);
1395 folio_start_writeback(folio);
1396 folio_unlock(folio);
1399 * Preserve the original error if there was one; catch
1400 * submission errors here and propagate into subsequent ioend
1403 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1406 list_del_init(&ioend->io_list);
1407 error2 = iomap_submit_ioend(wpc, ioend, error);
1408 if (error2 && !error)
1413 * We can end up here with no error and nothing to write only if we race
1414 * with a partial page truncate on a sub-page block sized filesystem.
1417 folio_end_writeback(folio);
1419 mapping_set_error(folio->mapping, error);
1424 * Write out a dirty page.
1426 * For delalloc space on the page, we need to allocate space and flush it.
1427 * For unwritten space on the page, we need to start the conversion to
1428 * regular allocated space.
1431 iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
1433 struct folio *folio = page_folio(page);
1434 struct iomap_writepage_ctx *wpc = data;
1435 struct inode *inode = folio->mapping->host;
1438 trace_iomap_writepage(inode, folio_pos(folio), folio_size(folio));
1441 * Refuse to write the folio out if we're called from reclaim context.
1443 * This avoids stack overflows when called from deeply used stacks in
1444 * random callers for direct reclaim or memcg reclaim. We explicitly
1445 * allow reclaim from kswapd as the stack usage there is relatively low.
1447 * This should never happen except in the case of a VM regression so
1450 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1455 * Is this folio beyond the end of the file?
1457 * The folio index is less than the end_index, adjust the end_pos
1458 * to the highest offset that this folio should represent.
1459 * -----------------------------------------------------
1460 * | file mapping | <EOF> |
1461 * -----------------------------------------------------
1462 * | Page ... | Page N-2 | Page N-1 | Page N | |
1463 * ^--------------------------------^----------|--------
1464 * | desired writeback range | see else |
1465 * ---------------------------------^------------------|
1467 isize = i_size_read(inode);
1468 end_pos = folio_pos(folio) + folio_size(folio);
1469 if (end_pos > isize) {
1471 * Check whether the page to write out is beyond or straddles
1473 * -------------------------------------------------------
1474 * | file mapping | <EOF> |
1475 * -------------------------------------------------------
1476 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1477 * ^--------------------------------^-----------|---------
1479 * ---------------------------------^-----------|--------|
1481 size_t poff = offset_in_folio(folio, isize);
1482 pgoff_t end_index = isize >> PAGE_SHIFT;
1485 * Skip the page if it's fully outside i_size, e.g. due to a
1486 * truncate operation that's in progress. We must redirty the
1487 * page so that reclaim stops reclaiming it. Otherwise
1488 * iomap_vm_releasepage() is called on it and gets confused.
1490 * Note that the end_index is unsigned long. If the given
1491 * offset is greater than 16TB on a 32-bit system then if we
1492 * checked if the page is fully outside i_size with
1493 * "if (page->index >= end_index + 1)", "end_index + 1" would
1494 * overflow and evaluate to 0. Hence this page would be
1495 * redirtied and written out repeatedly, which would result in
1496 * an infinite loop; the user program performing this operation
1497 * would hang. Instead, we can detect this situation by
1498 * checking if the page is totally beyond i_size or if its
1499 * offset is just equal to the EOF.
1501 if (folio->index > end_index ||
1502 (folio->index == end_index && poff == 0))
1506 * The page straddles i_size. It must be zeroed out on each
1507 * and every writepage invocation because it may be mmapped.
1508 * "A file is mapped in multiples of the page size. For a file
1509 * that is not a multiple of the page size, the remaining
1510 * memory is zeroed when mapped, and writes to that region are
1511 * not written out to the file."
1513 folio_zero_segment(folio, poff, folio_size(folio));
1517 return iomap_writepage_map(wpc, wbc, inode, folio, end_pos);
1520 folio_redirty_for_writepage(wbc, folio);
1521 folio_unlock(folio);
1526 iomap_writepage(struct page *page, struct writeback_control *wbc,
1527 struct iomap_writepage_ctx *wpc,
1528 const struct iomap_writeback_ops *ops)
1533 ret = iomap_do_writepage(page, wbc, wpc);
1536 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1538 EXPORT_SYMBOL_GPL(iomap_writepage);
1541 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1542 struct iomap_writepage_ctx *wpc,
1543 const struct iomap_writeback_ops *ops)
1548 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1551 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1553 EXPORT_SYMBOL_GPL(iomap_writepages);
1555 static int __init iomap_init(void)
1557 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
1558 offsetof(struct iomap_ioend, io_inline_bio),
1561 fs_initcall(iomap_init);