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;
439 unsigned first, last, i;
444 /* Limit range to this folio */
445 len = min(folio_size(folio) - from, count);
447 /* First and last blocks in range within page */
448 first = from >> inode->i_blkbits;
449 last = (from + len - 1) >> inode->i_blkbits;
451 for (i = first; i <= last; i++)
452 if (!test_bit(i, iop->uptodate))
456 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
459 iomap_releasepage(struct page *page, gfp_t gfp_mask)
461 struct folio *folio = page_folio(page);
463 trace_iomap_releasepage(folio->mapping->host, folio_pos(folio),
467 * mm accommodates an old ext3 case where clean pages might not have had
468 * the dirty bit cleared. Thus, it can send actual dirty pages to
469 * ->releasepage() via shrink_active_list(); skip those here.
471 if (folio_test_dirty(folio) || folio_test_writeback(folio))
473 iomap_page_release(folio);
476 EXPORT_SYMBOL_GPL(iomap_releasepage);
478 void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
480 trace_iomap_invalidate_folio(folio->mapping->host,
481 folio_pos(folio) + offset, len);
484 * If we're invalidating the entire folio, clear the dirty state
485 * from it and release it to avoid unnecessary buildup of the LRU.
487 if (offset == 0 && len == folio_size(folio)) {
488 WARN_ON_ONCE(folio_test_writeback(folio));
489 folio_cancel_dirty(folio);
490 iomap_page_release(folio);
491 } else if (folio_test_large(folio)) {
492 /* Must release the iop so the page can be split */
493 WARN_ON_ONCE(!folio_test_uptodate(folio) &&
494 folio_test_dirty(folio));
495 iomap_page_release(folio);
498 EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
500 #ifdef CONFIG_MIGRATION
502 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
503 struct page *page, enum migrate_mode mode)
505 struct folio *folio = page_folio(page);
506 struct folio *newfolio = page_folio(newpage);
509 ret = folio_migrate_mapping(mapping, newfolio, folio, 0);
510 if (ret != MIGRATEPAGE_SUCCESS)
513 if (folio_test_private(folio))
514 folio_attach_private(newfolio, folio_detach_private(folio));
516 if (mode != MIGRATE_SYNC_NO_COPY)
517 folio_migrate_copy(newfolio, folio);
519 folio_migrate_flags(newfolio, folio);
520 return MIGRATEPAGE_SUCCESS;
522 EXPORT_SYMBOL_GPL(iomap_migrate_page);
523 #endif /* CONFIG_MIGRATION */
526 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
528 loff_t i_size = i_size_read(inode);
531 * Only truncate newly allocated pages beyoned EOF, even if the
532 * write started inside the existing inode size.
534 if (pos + len > i_size)
535 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
538 static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
539 size_t poff, size_t plen, const struct iomap *iomap)
544 bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
545 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
546 bio_add_folio(&bio, folio, plen, poff);
547 return submit_bio_wait(&bio);
550 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
551 size_t len, struct folio *folio)
553 const struct iomap *srcmap = iomap_iter_srcmap(iter);
554 struct iomap_page *iop = iomap_page_create(iter->inode, folio);
555 loff_t block_size = i_blocksize(iter->inode);
556 loff_t block_start = round_down(pos, block_size);
557 loff_t block_end = round_up(pos + len, block_size);
558 size_t from = offset_in_folio(folio, pos), to = from + len;
561 if (folio_test_uptodate(folio))
563 folio_clear_error(folio);
566 iomap_adjust_read_range(iter->inode, folio, &block_start,
567 block_end - block_start, &poff, &plen);
571 if (!(iter->flags & IOMAP_UNSHARE) &&
572 (from <= poff || from >= poff + plen) &&
573 (to <= poff || to >= poff + plen))
576 if (iomap_block_needs_zeroing(iter, block_start)) {
577 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
579 folio_zero_segments(folio, poff, from, to, poff + plen);
581 int status = iomap_read_folio_sync(block_start, folio,
586 iomap_set_range_uptodate(folio, iop, poff, plen);
587 } while ((block_start += plen) < block_end);
592 static int iomap_write_begin_inline(const struct iomap_iter *iter,
595 /* needs more work for the tailpacking case; disable for now */
596 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
598 return iomap_read_inline_data(iter, folio);
601 static int iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
602 size_t len, struct folio **foliop)
604 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
605 const struct iomap *srcmap = iomap_iter_srcmap(iter);
607 unsigned fgp = FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE | FGP_NOFS;
610 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
611 if (srcmap != &iter->iomap)
612 BUG_ON(pos + len > srcmap->offset + srcmap->length);
614 if (fatal_signal_pending(current))
617 if (!mapping_large_folio_support(iter->inode->i_mapping))
618 len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
620 if (page_ops && page_ops->page_prepare) {
621 status = page_ops->page_prepare(iter->inode, pos, len);
626 folio = __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
627 fgp, mapping_gfp_mask(iter->inode->i_mapping));
632 if (pos + len > folio_pos(folio) + folio_size(folio))
633 len = folio_pos(folio) + folio_size(folio) - pos;
635 if (srcmap->type == IOMAP_INLINE)
636 status = iomap_write_begin_inline(iter, folio);
637 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
638 status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
640 status = __iomap_write_begin(iter, pos, len, folio);
642 if (unlikely(status))
651 iomap_write_failed(iter->inode, pos, len);
654 if (page_ops && page_ops->page_done)
655 page_ops->page_done(iter->inode, pos, 0, NULL);
659 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
660 size_t copied, struct folio *folio)
662 struct iomap_page *iop = to_iomap_page(folio);
663 flush_dcache_folio(folio);
666 * The blocks that were entirely written will now be uptodate, so we
667 * don't have to worry about a readpage reading them and overwriting a
668 * partial write. However, if we've encountered a short write and only
669 * partially written into a block, it will not be marked uptodate, so a
670 * readpage might come in and destroy our partial write.
672 * Do the simplest thing and just treat any short write to a
673 * non-uptodate page as a zero-length write, and force the caller to
674 * redo the whole thing.
676 if (unlikely(copied < len && !folio_test_uptodate(folio)))
678 iomap_set_range_uptodate(folio, iop, offset_in_folio(folio, pos), len);
679 filemap_dirty_folio(inode->i_mapping, folio);
683 static size_t iomap_write_end_inline(const struct iomap_iter *iter,
684 struct folio *folio, loff_t pos, size_t copied)
686 const struct iomap *iomap = &iter->iomap;
689 WARN_ON_ONCE(!folio_test_uptodate(folio));
690 BUG_ON(!iomap_inline_data_valid(iomap));
692 flush_dcache_folio(folio);
693 addr = kmap_local_folio(folio, pos);
694 memcpy(iomap_inline_data(iomap, pos), addr, copied);
697 mark_inode_dirty(iter->inode);
701 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
702 static size_t iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
703 size_t copied, struct folio *folio)
705 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
706 const struct iomap *srcmap = iomap_iter_srcmap(iter);
707 loff_t old_size = iter->inode->i_size;
710 if (srcmap->type == IOMAP_INLINE) {
711 ret = iomap_write_end_inline(iter, folio, pos, copied);
712 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
713 ret = block_write_end(NULL, iter->inode->i_mapping, pos, len,
714 copied, &folio->page, NULL);
716 ret = __iomap_write_end(iter->inode, pos, len, copied, folio);
720 * Update the in-memory inode size after copying the data into the page
721 * cache. It's up to the file system to write the updated size to disk,
722 * preferably after I/O completion so that no stale data is exposed.
724 if (pos + ret > old_size) {
725 i_size_write(iter->inode, pos + ret);
726 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
731 pagecache_isize_extended(iter->inode, old_size, pos);
732 if (page_ops && page_ops->page_done)
733 page_ops->page_done(iter->inode, pos, ret, &folio->page);
737 iomap_write_failed(iter->inode, pos, len);
741 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
743 loff_t length = iomap_length(iter);
744 loff_t pos = iter->pos;
751 unsigned long offset; /* Offset into pagecache page */
752 unsigned long bytes; /* Bytes to write to page */
753 size_t copied; /* Bytes copied from user */
755 offset = offset_in_page(pos);
756 bytes = min_t(unsigned long, PAGE_SIZE - offset,
763 * Bring in the user page that we'll copy from _first_.
764 * Otherwise there's a nasty deadlock on copying from the
765 * same page as we're writing to, without it being marked
768 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
773 status = iomap_write_begin(iter, pos, bytes, &folio);
774 if (unlikely(status))
777 page = folio_file_page(folio, pos >> PAGE_SHIFT);
778 if (mapping_writably_mapped(iter->inode->i_mapping))
779 flush_dcache_page(page);
781 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
783 status = iomap_write_end(iter, pos, bytes, copied, folio);
785 if (unlikely(copied != status))
786 iov_iter_revert(i, copied - status);
789 if (unlikely(status == 0)) {
791 * A short copy made iomap_write_end() reject the
792 * thing entirely. Might be memory poisoning
793 * halfway through, might be a race with munmap,
794 * might be severe memory pressure.
804 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
805 } while (iov_iter_count(i) && length);
807 return written ? written : status;
811 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
812 const struct iomap_ops *ops)
814 struct iomap_iter iter = {
815 .inode = iocb->ki_filp->f_mapping->host,
817 .len = iov_iter_count(i),
818 .flags = IOMAP_WRITE,
822 while ((ret = iomap_iter(&iter, ops)) > 0)
823 iter.processed = iomap_write_iter(&iter, i);
824 if (iter.pos == iocb->ki_pos)
826 return iter.pos - iocb->ki_pos;
828 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
830 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
832 struct iomap *iomap = &iter->iomap;
833 const struct iomap *srcmap = iomap_iter_srcmap(iter);
834 loff_t pos = iter->pos;
835 loff_t length = iomap_length(iter);
839 /* don't bother with blocks that are not shared to start with */
840 if (!(iomap->flags & IOMAP_F_SHARED))
842 /* don't bother with holes or unwritten extents */
843 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
847 unsigned long offset = offset_in_page(pos);
848 unsigned long bytes = min_t(loff_t, PAGE_SIZE - offset, length);
851 status = iomap_write_begin(iter, pos, bytes, &folio);
852 if (unlikely(status))
855 status = iomap_write_end(iter, pos, bytes, bytes, folio);
856 if (WARN_ON_ONCE(status == 0))
865 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
872 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
873 const struct iomap_ops *ops)
875 struct iomap_iter iter = {
879 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
883 while ((ret = iomap_iter(&iter, ops)) > 0)
884 iter.processed = iomap_unshare_iter(&iter);
887 EXPORT_SYMBOL_GPL(iomap_file_unshare);
889 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
891 const struct iomap *srcmap = iomap_iter_srcmap(iter);
892 loff_t pos = iter->pos;
893 loff_t length = iomap_length(iter);
896 /* already zeroed? we're done. */
897 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
904 size_t bytes = min_t(u64, SIZE_MAX, length);
906 status = iomap_write_begin(iter, pos, bytes, &folio);
910 offset = offset_in_folio(folio, pos);
911 if (bytes > folio_size(folio) - offset)
912 bytes = folio_size(folio) - offset;
914 folio_zero_range(folio, offset, bytes);
915 folio_mark_accessed(folio);
917 bytes = iomap_write_end(iter, pos, bytes, bytes, folio);
918 if (WARN_ON_ONCE(bytes == 0))
926 } while (length > 0);
932 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
933 const struct iomap_ops *ops)
935 struct iomap_iter iter = {
943 while ((ret = iomap_iter(&iter, ops)) > 0)
944 iter.processed = iomap_zero_iter(&iter, did_zero);
947 EXPORT_SYMBOL_GPL(iomap_zero_range);
950 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
951 const struct iomap_ops *ops)
953 unsigned int blocksize = i_blocksize(inode);
954 unsigned int off = pos & (blocksize - 1);
956 /* Block boundary? Nothing to do */
959 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
961 EXPORT_SYMBOL_GPL(iomap_truncate_page);
963 static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
966 loff_t length = iomap_length(iter);
969 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
970 ret = __block_write_begin_int(folio, iter->pos, length, NULL,
974 block_commit_write(&folio->page, 0, length);
976 WARN_ON_ONCE(!folio_test_uptodate(folio));
977 folio_mark_dirty(folio);
983 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
985 struct iomap_iter iter = {
986 .inode = file_inode(vmf->vma->vm_file),
987 .flags = IOMAP_WRITE | IOMAP_FAULT,
989 struct folio *folio = page_folio(vmf->page);
993 ret = folio_mkwrite_check_truncate(folio, iter.inode);
996 iter.pos = folio_pos(folio);
998 while ((ret = iomap_iter(&iter, ops)) > 0)
999 iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1003 folio_wait_stable(folio);
1004 return VM_FAULT_LOCKED;
1006 folio_unlock(folio);
1007 return block_page_mkwrite_return(ret);
1009 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1011 static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1012 size_t len, int error)
1014 struct iomap_page *iop = to_iomap_page(folio);
1017 folio_set_error(folio);
1018 mapping_set_error(inode->i_mapping, error);
1021 WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !iop);
1022 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) <= 0);
1024 if (!iop || atomic_sub_and_test(len, &iop->write_bytes_pending))
1025 folio_end_writeback(folio);
1029 * We're now finished for good with this ioend structure. Update the page
1030 * state, release holds on bios, and finally free up memory. Do not use the
1034 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1036 struct inode *inode = ioend->io_inode;
1037 struct bio *bio = &ioend->io_inline_bio;
1038 struct bio *last = ioend->io_bio, *next;
1039 u64 start = bio->bi_iter.bi_sector;
1040 loff_t offset = ioend->io_offset;
1041 bool quiet = bio_flagged(bio, BIO_QUIET);
1042 u32 folio_count = 0;
1044 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1045 struct folio_iter fi;
1048 * For the last bio, bi_private points to the ioend, so we
1049 * need to explicitly end the iteration here.
1054 next = bio->bi_private;
1056 /* walk all folios in bio, ending page IO on them */
1057 bio_for_each_folio_all(fi, bio) {
1058 iomap_finish_folio_write(inode, fi.folio, fi.length,
1064 /* The ioend has been freed by bio_put() */
1066 if (unlikely(error && !quiet)) {
1067 printk_ratelimited(KERN_ERR
1068 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1069 inode->i_sb->s_id, inode->i_ino, offset, start);
1075 * Ioend completion routine for merged bios. This can only be called from task
1076 * contexts as merged ioends can be of unbound length. Hence we have to break up
1077 * the writeback completions into manageable chunks to avoid long scheduler
1078 * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1079 * good batch processing throughput without creating adverse scheduler latency
1083 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1085 struct list_head tmp;
1090 list_replace_init(&ioend->io_list, &tmp);
1091 completions = iomap_finish_ioend(ioend, error);
1093 while (!list_empty(&tmp)) {
1094 if (completions > IOEND_BATCH_SIZE * 8) {
1098 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1099 list_del_init(&ioend->io_list);
1100 completions += iomap_finish_ioend(ioend, error);
1103 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1106 * We can merge two adjacent ioends if they have the same set of work to do.
1109 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1111 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1113 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1114 (next->io_flags & IOMAP_F_SHARED))
1116 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1117 (next->io_type == IOMAP_UNWRITTEN))
1119 if (ioend->io_offset + ioend->io_size != next->io_offset)
1122 * Do not merge physically discontiguous ioends. The filesystem
1123 * completion functions will have to iterate the physical
1124 * discontiguities even if we merge the ioends at a logical level, so
1125 * we don't gain anything by merging physical discontiguities here.
1127 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1128 * submission so does not point to the start sector of the bio at
1131 if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1137 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1139 struct iomap_ioend *next;
1141 INIT_LIST_HEAD(&ioend->io_list);
1143 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1145 if (!iomap_ioend_can_merge(ioend, next))
1147 list_move_tail(&next->io_list, &ioend->io_list);
1148 ioend->io_size += next->io_size;
1151 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1154 iomap_ioend_compare(void *priv, const struct list_head *a,
1155 const struct list_head *b)
1157 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1158 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1160 if (ia->io_offset < ib->io_offset)
1162 if (ia->io_offset > ib->io_offset)
1168 iomap_sort_ioends(struct list_head *ioend_list)
1170 list_sort(NULL, ioend_list, iomap_ioend_compare);
1172 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1174 static void iomap_writepage_end_bio(struct bio *bio)
1176 struct iomap_ioend *ioend = bio->bi_private;
1178 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1182 * Submit the final bio for an ioend.
1184 * If @error is non-zero, it means that we have a situation where some part of
1185 * the submission process has failed after we've marked pages for writeback
1186 * and unlocked them. In this situation, we need to fail the bio instead of
1187 * submitting it. This typically only happens on a filesystem shutdown.
1190 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1193 ioend->io_bio->bi_private = ioend;
1194 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1196 if (wpc->ops->prepare_ioend)
1197 error = wpc->ops->prepare_ioend(ioend, error);
1200 * If we're failing the IO now, just mark the ioend with an
1201 * error and finish it. This will run IO completion immediately
1202 * as there is only one reference to the ioend at this point in
1205 ioend->io_bio->bi_status = errno_to_blk_status(error);
1206 bio_endio(ioend->io_bio);
1210 submit_bio(ioend->io_bio);
1214 static struct iomap_ioend *
1215 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1216 loff_t offset, sector_t sector, struct writeback_control *wbc)
1218 struct iomap_ioend *ioend;
1221 bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1222 REQ_OP_WRITE | wbc_to_write_flags(wbc),
1223 GFP_NOFS, &iomap_ioend_bioset);
1224 bio->bi_iter.bi_sector = sector;
1225 wbc_init_bio(wbc, bio);
1227 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1228 INIT_LIST_HEAD(&ioend->io_list);
1229 ioend->io_type = wpc->iomap.type;
1230 ioend->io_flags = wpc->iomap.flags;
1231 ioend->io_inode = inode;
1233 ioend->io_folios = 0;
1234 ioend->io_offset = offset;
1235 ioend->io_bio = bio;
1236 ioend->io_sector = sector;
1241 * Allocate a new bio, and chain the old bio to the new one.
1243 * Note that we have to perform the chaining in this unintuitive order
1244 * so that the bi_private linkage is set up in the right direction for the
1245 * traversal in iomap_finish_ioend().
1248 iomap_chain_bio(struct bio *prev)
1252 new = bio_alloc(prev->bi_bdev, BIO_MAX_VECS, prev->bi_opf, GFP_NOFS);
1253 bio_clone_blkg_association(new, prev);
1254 new->bi_iter.bi_sector = bio_end_sector(prev);
1256 bio_chain(prev, new);
1257 bio_get(prev); /* for iomap_finish_ioend */
1263 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1266 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1267 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1269 if (wpc->iomap.type != wpc->ioend->io_type)
1271 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1273 if (sector != bio_end_sector(wpc->ioend->io_bio))
1276 * Limit ioend bio chain lengths to minimise IO completion latency. This
1277 * also prevents long tight loops ending page writeback on all the
1278 * folios in the ioend.
1280 if (wpc->ioend->io_folios >= IOEND_BATCH_SIZE)
1286 * Test to see if we have an existing ioend structure that we could append to
1287 * first; otherwise finish off the current ioend and start another.
1290 iomap_add_to_ioend(struct inode *inode, loff_t pos, struct folio *folio,
1291 struct iomap_page *iop, struct iomap_writepage_ctx *wpc,
1292 struct writeback_control *wbc, struct list_head *iolist)
1294 sector_t sector = iomap_sector(&wpc->iomap, pos);
1295 unsigned len = i_blocksize(inode);
1296 size_t poff = offset_in_folio(folio, pos);
1298 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos, sector)) {
1300 list_add(&wpc->ioend->io_list, iolist);
1301 wpc->ioend = iomap_alloc_ioend(inode, wpc, pos, sector, wbc);
1304 if (!bio_add_folio(wpc->ioend->io_bio, folio, len, poff)) {
1305 wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1306 bio_add_folio(wpc->ioend->io_bio, folio, len, poff);
1310 atomic_add(len, &iop->write_bytes_pending);
1311 wpc->ioend->io_size += len;
1312 wbc_account_cgroup_owner(wbc, &folio->page, len);
1316 * We implement an immediate ioend submission policy here to avoid needing to
1317 * chain multiple ioends and hence nest mempool allocations which can violate
1318 * the forward progress guarantees we need to provide. The current ioend we're
1319 * adding blocks to is cached in the writepage context, and if the new block
1320 * doesn't append to the cached ioend, it will create a new ioend and cache that
1323 * If a new ioend is created and cached, the old ioend is returned and queued
1324 * locally for submission once the entire page is processed or an error has been
1325 * detected. While ioends are submitted immediately after they are completed,
1326 * batching optimisations are provided by higher level block plugging.
1328 * At the end of a writeback pass, there will be a cached ioend remaining on the
1329 * writepage context that the caller will need to submit.
1332 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1333 struct writeback_control *wbc, struct inode *inode,
1334 struct folio *folio, u64 end_pos)
1336 struct iomap_page *iop = iomap_page_create(inode, folio);
1337 struct iomap_ioend *ioend, *next;
1338 unsigned len = i_blocksize(inode);
1339 unsigned nblocks = i_blocks_per_folio(inode, folio);
1340 u64 pos = folio_pos(folio);
1341 int error = 0, count = 0, i;
1342 LIST_HEAD(submit_list);
1344 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) != 0);
1347 * Walk through the folio to find areas to write back. If we
1348 * run off the end of the current map or find the current map
1349 * invalid, grab a new one.
1351 for (i = 0; i < nblocks && pos < end_pos; i++, pos += len) {
1352 if (iop && !test_bit(i, iop->uptodate))
1355 error = wpc->ops->map_blocks(wpc, inode, pos);
1358 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1360 if (wpc->iomap.type == IOMAP_HOLE)
1362 iomap_add_to_ioend(inode, pos, folio, iop, wpc, wbc,
1367 wpc->ioend->io_folios++;
1369 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1370 WARN_ON_ONCE(!folio_test_locked(folio));
1371 WARN_ON_ONCE(folio_test_writeback(folio));
1372 WARN_ON_ONCE(folio_test_dirty(folio));
1375 * We cannot cancel the ioend directly here on error. We may have
1376 * already set other pages under writeback and hence we have to run I/O
1377 * completion to mark the error state of the pages under writeback
1380 if (unlikely(error)) {
1382 * Let the filesystem know what portion of the current page
1383 * failed to map. If the page hasn't been added to ioend, it
1384 * won't be affected by I/O completion and we must unlock it
1387 if (wpc->ops->discard_folio)
1388 wpc->ops->discard_folio(folio, pos);
1390 folio_clear_uptodate(folio);
1391 folio_unlock(folio);
1396 folio_start_writeback(folio);
1397 folio_unlock(folio);
1400 * Preserve the original error if there was one; catch
1401 * submission errors here and propagate into subsequent ioend
1404 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1407 list_del_init(&ioend->io_list);
1408 error2 = iomap_submit_ioend(wpc, ioend, error);
1409 if (error2 && !error)
1414 * We can end up here with no error and nothing to write only if we race
1415 * with a partial page truncate on a sub-page block sized filesystem.
1418 folio_end_writeback(folio);
1420 mapping_set_error(folio->mapping, error);
1425 * Write out a dirty page.
1427 * For delalloc space on the page, we need to allocate space and flush it.
1428 * For unwritten space on the page, we need to start the conversion to
1429 * regular allocated space.
1432 iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
1434 struct folio *folio = page_folio(page);
1435 struct iomap_writepage_ctx *wpc = data;
1436 struct inode *inode = folio->mapping->host;
1439 trace_iomap_writepage(inode, folio_pos(folio), folio_size(folio));
1442 * Refuse to write the folio out if we're called from reclaim context.
1444 * This avoids stack overflows when called from deeply used stacks in
1445 * random callers for direct reclaim or memcg reclaim. We explicitly
1446 * allow reclaim from kswapd as the stack usage there is relatively low.
1448 * This should never happen except in the case of a VM regression so
1451 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1456 * Is this folio beyond the end of the file?
1458 * The folio index is less than the end_index, adjust the end_pos
1459 * to the highest offset that this folio should represent.
1460 * -----------------------------------------------------
1461 * | file mapping | <EOF> |
1462 * -----------------------------------------------------
1463 * | Page ... | Page N-2 | Page N-1 | Page N | |
1464 * ^--------------------------------^----------|--------
1465 * | desired writeback range | see else |
1466 * ---------------------------------^------------------|
1468 isize = i_size_read(inode);
1469 end_pos = folio_pos(folio) + folio_size(folio);
1470 if (end_pos > isize) {
1472 * Check whether the page to write out is beyond or straddles
1474 * -------------------------------------------------------
1475 * | file mapping | <EOF> |
1476 * -------------------------------------------------------
1477 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1478 * ^--------------------------------^-----------|---------
1480 * ---------------------------------^-----------|--------|
1482 size_t poff = offset_in_folio(folio, isize);
1483 pgoff_t end_index = isize >> PAGE_SHIFT;
1486 * Skip the page if it's fully outside i_size, e.g. due to a
1487 * truncate operation that's in progress. We must redirty the
1488 * page so that reclaim stops reclaiming it. Otherwise
1489 * iomap_vm_releasepage() is called on it and gets confused.
1491 * Note that the end_index is unsigned long. If the given
1492 * offset is greater than 16TB on a 32-bit system then if we
1493 * checked if the page is fully outside i_size with
1494 * "if (page->index >= end_index + 1)", "end_index + 1" would
1495 * overflow and evaluate to 0. Hence this page would be
1496 * redirtied and written out repeatedly, which would result in
1497 * an infinite loop; the user program performing this operation
1498 * would hang. Instead, we can detect this situation by
1499 * checking if the page is totally beyond i_size or if its
1500 * offset is just equal to the EOF.
1502 if (folio->index > end_index ||
1503 (folio->index == end_index && poff == 0))
1507 * The page straddles i_size. It must be zeroed out on each
1508 * and every writepage invocation because it may be mmapped.
1509 * "A file is mapped in multiples of the page size. For a file
1510 * that is not a multiple of the page size, the remaining
1511 * memory is zeroed when mapped, and writes to that region are
1512 * not written out to the file."
1514 folio_zero_segment(folio, poff, folio_size(folio));
1518 return iomap_writepage_map(wpc, wbc, inode, folio, end_pos);
1521 folio_redirty_for_writepage(wbc, folio);
1522 folio_unlock(folio);
1527 iomap_writepage(struct page *page, struct writeback_control *wbc,
1528 struct iomap_writepage_ctx *wpc,
1529 const struct iomap_writeback_ops *ops)
1534 ret = iomap_do_writepage(page, wbc, wpc);
1537 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1539 EXPORT_SYMBOL_GPL(iomap_writepage);
1542 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1543 struct iomap_writepage_ctx *wpc,
1544 const struct iomap_writeback_ops *ops)
1549 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1552 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1554 EXPORT_SYMBOL_GPL(iomap_writepages);
1556 static int __init iomap_init(void)
1558 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
1559 offsetof(struct iomap_ioend, io_inline_bio),
1562 fs_initcall(iomap_init);