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"
25 * Structure allocated for each page or THP when block size < page size
26 * to track sub-page uptodate status and I/O completions.
29 atomic_t read_bytes_pending;
30 atomic_t write_bytes_pending;
31 spinlock_t uptodate_lock;
32 unsigned long uptodate[];
35 static inline struct iomap_page *to_iomap_page(struct page *page)
38 * per-block data is stored in the head page. Callers should
39 * not be dealing with tail pages (and if they are, they can
40 * call thp_head() first.
42 VM_BUG_ON_PGFLAGS(PageTail(page), page);
44 if (page_has_private(page))
45 return (struct iomap_page *)page_private(page);
49 static struct bio_set iomap_ioend_bioset;
51 static struct iomap_page *
52 iomap_page_create(struct inode *inode, struct page *page)
54 struct iomap_page *iop = to_iomap_page(page);
55 unsigned int nr_blocks = i_blocks_per_page(inode, page);
57 if (iop || nr_blocks <= 1)
60 iop = kzalloc(struct_size(iop, uptodate, BITS_TO_LONGS(nr_blocks)),
61 GFP_NOFS | __GFP_NOFAIL);
62 spin_lock_init(&iop->uptodate_lock);
63 if (PageUptodate(page))
64 bitmap_fill(iop->uptodate, nr_blocks);
65 attach_page_private(page, iop);
70 iomap_page_release(struct page *page)
72 struct iomap_page *iop = detach_page_private(page);
73 unsigned int nr_blocks = i_blocks_per_page(page->mapping->host, page);
77 WARN_ON_ONCE(atomic_read(&iop->read_bytes_pending));
78 WARN_ON_ONCE(atomic_read(&iop->write_bytes_pending));
79 WARN_ON_ONCE(bitmap_full(iop->uptodate, nr_blocks) !=
85 * Calculate the range inside the page that we actually need to read.
88 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
89 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
91 loff_t orig_pos = *pos;
92 loff_t isize = i_size_read(inode);
93 unsigned block_bits = inode->i_blkbits;
94 unsigned block_size = (1 << block_bits);
95 unsigned poff = offset_in_page(*pos);
96 unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
97 unsigned first = poff >> block_bits;
98 unsigned last = (poff + plen - 1) >> block_bits;
101 * If the block size is smaller than the page size we need to check the
102 * per-block uptodate status and adjust the offset and length if needed
103 * to avoid reading in already uptodate ranges.
108 /* move forward for each leading block marked uptodate */
109 for (i = first; i <= last; i++) {
110 if (!test_bit(i, iop->uptodate))
118 /* truncate len if we find any trailing uptodate block(s) */
119 for ( ; i <= last; i++) {
120 if (test_bit(i, iop->uptodate)) {
121 plen -= (last - i + 1) * block_size;
129 * If the extent spans the block that contains the i_size we need to
130 * handle both halves separately so that we properly zero data in the
131 * page cache for blocks that are entirely outside of i_size.
133 if (orig_pos <= isize && orig_pos + length > isize) {
134 unsigned end = offset_in_page(isize - 1) >> block_bits;
136 if (first <= end && last > end)
137 plen -= (last - end) * block_size;
145 iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
147 struct iomap_page *iop = to_iomap_page(page);
148 struct inode *inode = page->mapping->host;
149 unsigned first = off >> inode->i_blkbits;
150 unsigned last = (off + len - 1) >> inode->i_blkbits;
153 spin_lock_irqsave(&iop->uptodate_lock, flags);
154 bitmap_set(iop->uptodate, first, last - first + 1);
155 if (bitmap_full(iop->uptodate, i_blocks_per_page(inode, page)))
156 SetPageUptodate(page);
157 spin_unlock_irqrestore(&iop->uptodate_lock, flags);
161 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
166 if (page_has_private(page))
167 iomap_iop_set_range_uptodate(page, off, len);
169 SetPageUptodate(page);
173 iomap_read_page_end_io(struct bio_vec *bvec, int error)
175 struct page *page = bvec->bv_page;
176 struct iomap_page *iop = to_iomap_page(page);
178 if (unlikely(error)) {
179 ClearPageUptodate(page);
182 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
185 if (!iop || atomic_sub_and_test(bvec->bv_len, &iop->read_bytes_pending))
190 iomap_read_end_io(struct bio *bio)
192 int error = blk_status_to_errno(bio->bi_status);
193 struct bio_vec *bvec;
194 struct bvec_iter_all iter_all;
196 bio_for_each_segment_all(bvec, bio, iter_all)
197 iomap_read_page_end_io(bvec, error);
201 struct iomap_readpage_ctx {
202 struct page *cur_page;
203 bool cur_page_in_bio;
205 struct readahead_control *rac;
209 iomap_read_inline_data(struct inode *inode, struct page *page,
212 size_t size = i_size_read(inode);
215 if (PageUptodate(page))
219 BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
221 addr = kmap_atomic(page);
222 memcpy(addr, iomap->inline_data, size);
223 memset(addr + size, 0, PAGE_SIZE - size);
225 SetPageUptodate(page);
228 static inline bool iomap_block_needs_zeroing(struct inode *inode,
229 struct iomap *iomap, loff_t pos)
231 return iomap->type != IOMAP_MAPPED ||
232 (iomap->flags & IOMAP_F_NEW) ||
233 pos >= i_size_read(inode);
237 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
238 struct iomap *iomap, struct iomap *srcmap)
240 struct iomap_readpage_ctx *ctx = data;
241 struct page *page = ctx->cur_page;
242 struct iomap_page *iop = iomap_page_create(inode, page);
243 bool same_page = false, is_contig = false;
244 loff_t orig_pos = pos;
248 if (iomap->type == IOMAP_INLINE) {
250 iomap_read_inline_data(inode, page, iomap);
254 /* zero post-eof blocks as the page may be mapped */
255 iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
259 if (iomap_block_needs_zeroing(inode, iomap, pos)) {
260 zero_user(page, poff, plen);
261 iomap_set_range_uptodate(page, poff, plen);
265 ctx->cur_page_in_bio = true;
267 atomic_add(plen, &iop->read_bytes_pending);
269 /* Try to merge into a previous segment if we can */
270 sector = iomap_sector(iomap, pos);
271 if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
272 if (__bio_try_merge_page(ctx->bio, page, plen, poff,
278 if (!is_contig || bio_full(ctx->bio, plen)) {
279 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
280 gfp_t orig_gfp = gfp;
281 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
284 submit_bio(ctx->bio);
286 if (ctx->rac) /* same as readahead_gfp_mask */
287 gfp |= __GFP_NORETRY | __GFP_NOWARN;
288 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
290 * If the bio_alloc fails, try it again for a single page to
291 * avoid having to deal with partial page reads. This emulates
292 * what do_mpage_readpage does.
295 ctx->bio = bio_alloc(orig_gfp, 1);
296 ctx->bio->bi_opf = REQ_OP_READ;
298 ctx->bio->bi_opf |= REQ_RAHEAD;
299 ctx->bio->bi_iter.bi_sector = sector;
300 bio_set_dev(ctx->bio, iomap->bdev);
301 ctx->bio->bi_end_io = iomap_read_end_io;
304 bio_add_page(ctx->bio, page, plen, poff);
307 * Move the caller beyond our range so that it keeps making progress.
308 * For that we have to include any leading non-uptodate ranges, but
309 * we can skip trailing ones as they will be handled in the next
312 return pos - orig_pos + plen;
316 iomap_readpage(struct page *page, const struct iomap_ops *ops)
318 struct iomap_readpage_ctx ctx = { .cur_page = page };
319 struct inode *inode = page->mapping->host;
323 trace_iomap_readpage(page->mapping->host, 1);
325 for (poff = 0; poff < PAGE_SIZE; poff += ret) {
326 ret = iomap_apply(inode, page_offset(page) + poff,
327 PAGE_SIZE - poff, 0, ops, &ctx,
328 iomap_readpage_actor);
330 WARN_ON_ONCE(ret == 0);
338 WARN_ON_ONCE(!ctx.cur_page_in_bio);
340 WARN_ON_ONCE(ctx.cur_page_in_bio);
345 * Just like mpage_readahead and block_read_full_page we always
346 * return 0 and just mark the page as PageError on errors. This
347 * should be cleaned up all through the stack eventually.
351 EXPORT_SYMBOL_GPL(iomap_readpage);
354 iomap_readahead_actor(struct inode *inode, loff_t pos, loff_t length,
355 void *data, struct iomap *iomap, struct iomap *srcmap)
357 struct iomap_readpage_ctx *ctx = data;
360 for (done = 0; done < length; done += ret) {
361 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
362 if (!ctx->cur_page_in_bio)
363 unlock_page(ctx->cur_page);
364 put_page(ctx->cur_page);
365 ctx->cur_page = NULL;
367 if (!ctx->cur_page) {
368 ctx->cur_page = readahead_page(ctx->rac);
369 ctx->cur_page_in_bio = false;
371 ret = iomap_readpage_actor(inode, pos + done, length - done,
379 * iomap_readahead - Attempt to read pages from a file.
380 * @rac: Describes the pages to be read.
381 * @ops: The operations vector for the filesystem.
383 * This function is for filesystems to call to implement their readahead
384 * address_space operation.
386 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
387 * blocks from disc), and may wait for it. The caller may be trying to
388 * access a different page, and so sleeping excessively should be avoided.
389 * It may allocate memory, but should avoid costly allocations. This
390 * function is called with memalloc_nofs set, so allocations will not cause
391 * the filesystem to be reentered.
393 void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
395 struct inode *inode = rac->mapping->host;
396 loff_t pos = readahead_pos(rac);
397 loff_t length = readahead_length(rac);
398 struct iomap_readpage_ctx ctx = {
402 trace_iomap_readahead(inode, readahead_count(rac));
405 loff_t ret = iomap_apply(inode, pos, length, 0, ops,
406 &ctx, iomap_readahead_actor);
408 WARN_ON_ONCE(ret == 0);
418 if (!ctx.cur_page_in_bio)
419 unlock_page(ctx.cur_page);
420 put_page(ctx.cur_page);
423 EXPORT_SYMBOL_GPL(iomap_readahead);
426 * iomap_is_partially_uptodate checks whether blocks within a page are
429 * Returns true if all blocks which correspond to a file portion
430 * we want to read within the page are uptodate.
433 iomap_is_partially_uptodate(struct page *page, unsigned long from,
436 struct iomap_page *iop = to_iomap_page(page);
437 struct inode *inode = page->mapping->host;
438 unsigned len, first, last;
441 /* Limit range to one page */
442 len = min_t(unsigned, PAGE_SIZE - from, count);
444 /* First and last blocks in range within page */
445 first = from >> inode->i_blkbits;
446 last = (from + len - 1) >> inode->i_blkbits;
449 for (i = first; i <= last; i++)
450 if (!test_bit(i, iop->uptodate))
457 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
460 iomap_releasepage(struct page *page, gfp_t gfp_mask)
462 trace_iomap_releasepage(page->mapping->host, page_offset(page),
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 (PageDirty(page) || PageWriteback(page))
472 iomap_page_release(page);
475 EXPORT_SYMBOL_GPL(iomap_releasepage);
478 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
480 trace_iomap_invalidatepage(page->mapping->host, offset, len);
483 * If we are invalidating the entire page, clear the dirty state from it
484 * and release it to avoid unnecessary buildup of the LRU.
486 if (offset == 0 && len == PAGE_SIZE) {
487 WARN_ON_ONCE(PageWriteback(page));
488 cancel_dirty_page(page);
489 iomap_page_release(page);
492 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
494 #ifdef CONFIG_MIGRATION
496 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
497 struct page *page, enum migrate_mode mode)
501 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
502 if (ret != MIGRATEPAGE_SUCCESS)
505 if (page_has_private(page))
506 attach_page_private(newpage, detach_page_private(page));
508 if (mode != MIGRATE_SYNC_NO_COPY)
509 migrate_page_copy(newpage, page);
511 migrate_page_states(newpage, page);
512 return MIGRATEPAGE_SUCCESS;
514 EXPORT_SYMBOL_GPL(iomap_migrate_page);
515 #endif /* CONFIG_MIGRATION */
518 IOMAP_WRITE_F_UNSHARE = (1 << 0),
522 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
524 loff_t i_size = i_size_read(inode);
527 * Only truncate newly allocated pages beyoned EOF, even if the
528 * write started inside the existing inode size.
530 if (pos + len > i_size)
531 truncate_pagecache_range(inode, max(pos, i_size),
536 iomap_read_page_sync(loff_t block_start, struct page *page, unsigned poff,
537 unsigned plen, struct iomap *iomap)
542 bio_init(&bio, &bvec, 1);
543 bio.bi_opf = REQ_OP_READ;
544 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
545 bio_set_dev(&bio, iomap->bdev);
546 __bio_add_page(&bio, page, plen, poff);
547 return submit_bio_wait(&bio);
551 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, int flags,
552 struct page *page, struct iomap *srcmap)
554 struct iomap_page *iop = iomap_page_create(inode, page);
555 loff_t block_size = i_blocksize(inode);
556 loff_t block_start = round_down(pos, block_size);
557 loff_t block_end = round_up(pos + len, block_size);
558 unsigned from = offset_in_page(pos), to = from + len, poff, plen;
560 if (PageUptodate(page))
562 ClearPageError(page);
565 iomap_adjust_read_range(inode, iop, &block_start,
566 block_end - block_start, &poff, &plen);
570 if (!(flags & IOMAP_WRITE_F_UNSHARE) &&
571 (from <= poff || from >= poff + plen) &&
572 (to <= poff || to >= poff + plen))
575 if (iomap_block_needs_zeroing(inode, srcmap, block_start)) {
576 if (WARN_ON_ONCE(flags & IOMAP_WRITE_F_UNSHARE))
578 zero_user_segments(page, poff, from, to, poff + plen);
580 int status = iomap_read_page_sync(block_start, page,
585 iomap_set_range_uptodate(page, poff, plen);
586 } while ((block_start += plen) < block_end);
592 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
593 struct page **pagep, struct iomap *iomap, struct iomap *srcmap)
595 const struct iomap_page_ops *page_ops = iomap->page_ops;
599 BUG_ON(pos + len > iomap->offset + iomap->length);
601 BUG_ON(pos + len > srcmap->offset + srcmap->length);
603 if (fatal_signal_pending(current))
606 if (page_ops && page_ops->page_prepare) {
607 status = page_ops->page_prepare(inode, pos, len, iomap);
612 page = grab_cache_page_write_begin(inode->i_mapping, pos >> PAGE_SHIFT,
619 if (srcmap->type == IOMAP_INLINE)
620 iomap_read_inline_data(inode, page, srcmap);
621 else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
622 status = __block_write_begin_int(page, pos, len, NULL, srcmap);
624 status = __iomap_write_begin(inode, pos, len, flags, page,
627 if (unlikely(status))
636 iomap_write_failed(inode, pos, len);
639 if (page_ops && page_ops->page_done)
640 page_ops->page_done(inode, pos, 0, NULL, iomap);
645 iomap_set_page_dirty(struct page *page)
647 struct address_space *mapping = page_mapping(page);
650 if (unlikely(!mapping))
651 return !TestSetPageDirty(page);
654 * Lock out page->mem_cgroup migration to keep PageDirty
655 * synchronized with per-memcg dirty page counters.
657 lock_page_memcg(page);
658 newly_dirty = !TestSetPageDirty(page);
660 __set_page_dirty(page, mapping, 0);
661 unlock_page_memcg(page);
664 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
667 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
669 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
670 size_t copied, struct page *page)
672 flush_dcache_page(page);
675 * The blocks that were entirely written will now be uptodate, so we
676 * don't have to worry about a readpage reading them and overwriting a
677 * partial write. However if we have encountered a short write and only
678 * partially written into a block, it will not be marked uptodate, so a
679 * readpage might come in and destroy our partial write.
681 * Do the simplest thing, and just treat any short write to a non
682 * uptodate page as a zero-length write, and force the caller to redo
685 if (unlikely(copied < len && !PageUptodate(page)))
687 iomap_set_range_uptodate(page, offset_in_page(pos), len);
688 iomap_set_page_dirty(page);
692 static size_t iomap_write_end_inline(struct inode *inode, struct page *page,
693 struct iomap *iomap, loff_t pos, size_t copied)
697 WARN_ON_ONCE(!PageUptodate(page));
698 BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
700 flush_dcache_page(page);
701 addr = kmap_atomic(page);
702 memcpy(iomap->inline_data + pos, addr + pos, copied);
705 mark_inode_dirty(inode);
709 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
710 static size_t iomap_write_end(struct inode *inode, loff_t pos, size_t len,
711 size_t copied, struct page *page, struct iomap *iomap,
712 struct iomap *srcmap)
714 const struct iomap_page_ops *page_ops = iomap->page_ops;
715 loff_t old_size = inode->i_size;
718 if (srcmap->type == IOMAP_INLINE) {
719 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
720 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
721 ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
724 ret = __iomap_write_end(inode, pos, len, copied, page);
728 * Update the in-memory inode size after copying the data into the page
729 * cache. It's up to the file system to write the updated size to disk,
730 * preferably after I/O completion so that no stale data is exposed.
732 if (pos + ret > old_size) {
733 i_size_write(inode, pos + ret);
734 iomap->flags |= IOMAP_F_SIZE_CHANGED;
739 pagecache_isize_extended(inode, old_size, pos);
740 if (page_ops && page_ops->page_done)
741 page_ops->page_done(inode, pos, ret, page, iomap);
745 iomap_write_failed(inode, pos, len);
750 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
751 struct iomap *iomap, struct iomap *srcmap)
753 struct iov_iter *i = data;
759 unsigned long offset; /* Offset into pagecache page */
760 unsigned long bytes; /* Bytes to write to page */
761 size_t copied; /* Bytes copied from user */
763 offset = offset_in_page(pos);
764 bytes = min_t(unsigned long, PAGE_SIZE - offset,
771 * Bring in the user page that we will copy from _first_.
772 * Otherwise there's a nasty deadlock on copying from the
773 * same page as we're writing to, without it being marked
776 * Not only is this an optimisation, but it is also required
777 * to check that the address is actually valid, when atomic
778 * usercopies are used, below.
780 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
785 status = iomap_write_begin(inode, pos, bytes, 0, &page, iomap,
787 if (unlikely(status))
790 if (mapping_writably_mapped(inode->i_mapping))
791 flush_dcache_page(page);
793 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
795 copied = iomap_write_end(inode, pos, bytes, copied, page, iomap,
800 iov_iter_advance(i, copied);
801 if (unlikely(copied == 0)) {
803 * If we were unable to copy any data at all, we must
804 * fall back to a single segment length write.
806 * If we didn't fallback here, we could livelock
807 * because not all segments in the iov can be copied at
808 * once without a pagefault.
810 bytes = min_t(unsigned long, PAGE_SIZE - offset,
811 iov_iter_single_seg_count(i));
818 balance_dirty_pages_ratelimited(inode->i_mapping);
819 } while (iov_iter_count(i) && length);
821 return written ? written : status;
825 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
826 const struct iomap_ops *ops)
828 struct inode *inode = iocb->ki_filp->f_mapping->host;
829 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
831 while (iov_iter_count(iter)) {
832 ret = iomap_apply(inode, pos, iov_iter_count(iter),
833 IOMAP_WRITE, ops, iter, iomap_write_actor);
840 return written ? written : ret;
842 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
845 iomap_unshare_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
846 struct iomap *iomap, struct iomap *srcmap)
851 /* don't bother with blocks that are not shared to start with */
852 if (!(iomap->flags & IOMAP_F_SHARED))
854 /* don't bother with holes or unwritten extents */
855 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
859 unsigned long offset = offset_in_page(pos);
860 unsigned long bytes = min_t(loff_t, PAGE_SIZE - offset, length);
863 status = iomap_write_begin(inode, pos, bytes,
864 IOMAP_WRITE_F_UNSHARE, &page, iomap, srcmap);
865 if (unlikely(status))
868 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap,
870 if (WARN_ON_ONCE(status == 0))
879 balance_dirty_pages_ratelimited(inode->i_mapping);
886 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
887 const struct iomap_ops *ops)
892 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
893 iomap_unshare_actor);
902 EXPORT_SYMBOL_GPL(iomap_file_unshare);
904 static s64 iomap_zero(struct inode *inode, loff_t pos, u64 length,
905 struct iomap *iomap, struct iomap *srcmap)
909 unsigned offset = offset_in_page(pos);
910 unsigned bytes = min_t(u64, PAGE_SIZE - offset, length);
912 status = iomap_write_begin(inode, pos, bytes, 0, &page, iomap, srcmap);
916 zero_user(page, offset, bytes);
917 mark_page_accessed(page);
919 return iomap_write_end(inode, pos, bytes, bytes, page, iomap, srcmap);
922 static loff_t iomap_zero_range_actor(struct inode *inode, loff_t pos,
923 loff_t length, void *data, struct iomap *iomap,
924 struct iomap *srcmap)
926 bool *did_zero = data;
929 /* already zeroed? we're done. */
930 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
937 bytes = dax_iomap_zero(pos, length, iomap);
939 bytes = iomap_zero(inode, pos, length, iomap, srcmap);
948 } while (length > 0);
954 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
955 const struct iomap_ops *ops)
960 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
961 ops, did_zero, iomap_zero_range_actor);
971 EXPORT_SYMBOL_GPL(iomap_zero_range);
974 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
975 const struct iomap_ops *ops)
977 unsigned int blocksize = i_blocksize(inode);
978 unsigned int off = pos & (blocksize - 1);
980 /* Block boundary? Nothing to do */
983 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
985 EXPORT_SYMBOL_GPL(iomap_truncate_page);
988 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
989 void *data, struct iomap *iomap, struct iomap *srcmap)
991 struct page *page = data;
994 if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
995 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
998 block_commit_write(page, 0, length);
1000 WARN_ON_ONCE(!PageUptodate(page));
1001 iomap_page_create(inode, page);
1002 set_page_dirty(page);
1008 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1010 struct page *page = vmf->page;
1011 struct inode *inode = file_inode(vmf->vma->vm_file);
1012 unsigned long length;
1017 ret = page_mkwrite_check_truncate(page, inode);
1022 offset = page_offset(page);
1023 while (length > 0) {
1024 ret = iomap_apply(inode, offset, length,
1025 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1026 iomap_page_mkwrite_actor);
1027 if (unlikely(ret <= 0))
1033 wait_for_stable_page(page);
1034 return VM_FAULT_LOCKED;
1037 return block_page_mkwrite_return(ret);
1039 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1042 iomap_finish_page_writeback(struct inode *inode, struct page *page,
1043 int error, unsigned int len)
1045 struct iomap_page *iop = to_iomap_page(page);
1049 mapping_set_error(inode->i_mapping, error);
1052 WARN_ON_ONCE(i_blocks_per_page(inode, page) > 1 && !iop);
1053 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) <= 0);
1055 if (!iop || atomic_sub_and_test(len, &iop->write_bytes_pending))
1056 end_page_writeback(page);
1060 * We're now finished for good with this ioend structure. Update the page
1061 * state, release holds on bios, and finally free up memory. Do not use the
1065 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1067 struct inode *inode = ioend->io_inode;
1068 struct bio *bio = &ioend->io_inline_bio;
1069 struct bio *last = ioend->io_bio, *next;
1070 u64 start = bio->bi_iter.bi_sector;
1071 loff_t offset = ioend->io_offset;
1072 bool quiet = bio_flagged(bio, BIO_QUIET);
1074 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1076 struct bvec_iter_all iter_all;
1079 * For the last bio, bi_private points to the ioend, so we
1080 * need to explicitly end the iteration here.
1085 next = bio->bi_private;
1087 /* walk each page on bio, ending page IO on them */
1088 bio_for_each_segment_all(bv, bio, iter_all)
1089 iomap_finish_page_writeback(inode, bv->bv_page, error,
1093 /* The ioend has been freed by bio_put() */
1095 if (unlikely(error && !quiet)) {
1096 printk_ratelimited(KERN_ERR
1097 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1098 inode->i_sb->s_id, inode->i_ino, offset, start);
1103 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1105 struct list_head tmp;
1107 list_replace_init(&ioend->io_list, &tmp);
1108 iomap_finish_ioend(ioend, error);
1110 while (!list_empty(&tmp)) {
1111 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1112 list_del_init(&ioend->io_list);
1113 iomap_finish_ioend(ioend, error);
1116 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1119 * We can merge two adjacent ioends if they have the same set of work to do.
1122 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1124 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1126 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1127 (next->io_flags & IOMAP_F_SHARED))
1129 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1130 (next->io_type == IOMAP_UNWRITTEN))
1132 if (ioend->io_offset + ioend->io_size != next->io_offset)
1138 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends,
1139 void (*merge_private)(struct iomap_ioend *ioend,
1140 struct iomap_ioend *next))
1142 struct iomap_ioend *next;
1144 INIT_LIST_HEAD(&ioend->io_list);
1146 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1148 if (!iomap_ioend_can_merge(ioend, next))
1150 list_move_tail(&next->io_list, &ioend->io_list);
1151 ioend->io_size += next->io_size;
1152 if (next->io_private && merge_private)
1153 merge_private(ioend, next);
1156 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1159 iomap_ioend_compare(void *priv, const struct list_head *a,
1160 const struct list_head *b)
1162 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1163 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1165 if (ia->io_offset < ib->io_offset)
1167 if (ia->io_offset > ib->io_offset)
1173 iomap_sort_ioends(struct list_head *ioend_list)
1175 list_sort(NULL, ioend_list, iomap_ioend_compare);
1177 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1179 static void iomap_writepage_end_bio(struct bio *bio)
1181 struct iomap_ioend *ioend = bio->bi_private;
1183 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1187 * Submit the final bio for an ioend.
1189 * If @error is non-zero, it means that we have a situation where some part of
1190 * the submission process has failed after we have marked paged for writeback
1191 * and unlocked them. In this situation, we need to fail the bio instead of
1192 * submitting it. This typically only happens on a filesystem shutdown.
1195 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1198 ioend->io_bio->bi_private = ioend;
1199 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1201 if (wpc->ops->prepare_ioend)
1202 error = wpc->ops->prepare_ioend(ioend, error);
1205 * If we are failing the IO now, just mark the ioend with an
1206 * error and finish it. This will run IO completion immediately
1207 * as there is only one reference to the ioend at this point in
1210 ioend->io_bio->bi_status = errno_to_blk_status(error);
1211 bio_endio(ioend->io_bio);
1215 submit_bio(ioend->io_bio);
1219 static struct iomap_ioend *
1220 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1221 loff_t offset, sector_t sector, struct writeback_control *wbc)
1223 struct iomap_ioend *ioend;
1226 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &iomap_ioend_bioset);
1227 bio_set_dev(bio, wpc->iomap.bdev);
1228 bio->bi_iter.bi_sector = sector;
1229 bio->bi_opf = REQ_OP_WRITE | wbc_to_write_flags(wbc);
1230 bio->bi_write_hint = inode->i_write_hint;
1231 wbc_init_bio(wbc, bio);
1233 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1234 INIT_LIST_HEAD(&ioend->io_list);
1235 ioend->io_type = wpc->iomap.type;
1236 ioend->io_flags = wpc->iomap.flags;
1237 ioend->io_inode = inode;
1239 ioend->io_offset = offset;
1240 ioend->io_private = NULL;
1241 ioend->io_bio = bio;
1246 * Allocate a new bio, and chain the old bio to the new one.
1248 * Note that we have to do perform the chaining in this unintuitive order
1249 * so that the bi_private linkage is set up in the right direction for the
1250 * traversal in iomap_finish_ioend().
1253 iomap_chain_bio(struct bio *prev)
1257 new = bio_alloc(GFP_NOFS, BIO_MAX_PAGES);
1258 bio_copy_dev(new, prev);/* also copies over blkcg information */
1259 new->bi_iter.bi_sector = bio_end_sector(prev);
1260 new->bi_opf = prev->bi_opf;
1261 new->bi_write_hint = prev->bi_write_hint;
1263 bio_chain(prev, new);
1264 bio_get(prev); /* for iomap_finish_ioend */
1270 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1273 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1274 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1276 if (wpc->iomap.type != wpc->ioend->io_type)
1278 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1280 if (sector != bio_end_sector(wpc->ioend->io_bio))
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 offset, struct page *page,
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, offset);
1295 unsigned len = i_blocksize(inode);
1296 unsigned poff = offset & (PAGE_SIZE - 1);
1297 bool merged, same_page = false;
1299 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, offset, sector)) {
1301 list_add(&wpc->ioend->io_list, iolist);
1302 wpc->ioend = iomap_alloc_ioend(inode, wpc, offset, sector, wbc);
1305 merged = __bio_try_merge_page(wpc->ioend->io_bio, page, len, poff,
1308 atomic_add(len, &iop->write_bytes_pending);
1311 if (bio_full(wpc->ioend->io_bio, len)) {
1312 wpc->ioend->io_bio =
1313 iomap_chain_bio(wpc->ioend->io_bio);
1315 bio_add_page(wpc->ioend->io_bio, page, len, poff);
1318 wpc->ioend->io_size += len;
1319 wbc_account_cgroup_owner(wbc, page, len);
1323 * We implement an immediate ioend submission policy here to avoid needing to
1324 * chain multiple ioends and hence nest mempool allocations which can violate
1325 * forward progress guarantees we need to provide. The current ioend we are
1326 * adding blocks to is cached on the writepage context, and if the new block
1327 * does not append to the cached ioend it will create a new ioend and cache that
1330 * If a new ioend is created and cached, the old ioend is returned and queued
1331 * locally for submission once the entire page is processed or an error has been
1332 * detected. While ioends are submitted immediately after they are completed,
1333 * batching optimisations are provided by higher level block plugging.
1335 * At the end of a writeback pass, there will be a cached ioend remaining on the
1336 * writepage context that the caller will need to submit.
1339 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1340 struct writeback_control *wbc, struct inode *inode,
1341 struct page *page, u64 end_offset)
1343 struct iomap_page *iop = to_iomap_page(page);
1344 struct iomap_ioend *ioend, *next;
1345 unsigned len = i_blocksize(inode);
1346 u64 file_offset; /* file offset of page */
1347 int error = 0, count = 0, i;
1348 LIST_HEAD(submit_list);
1350 WARN_ON_ONCE(i_blocks_per_page(inode, page) > 1 && !iop);
1351 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) != 0);
1354 * Walk through the page to find areas to write back. If we run off the
1355 * end of the current map or find the current map invalid, grab a new
1358 for (i = 0, file_offset = page_offset(page);
1359 i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
1360 i++, file_offset += len) {
1361 if (iop && !test_bit(i, iop->uptodate))
1364 error = wpc->ops->map_blocks(wpc, inode, file_offset);
1367 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1369 if (wpc->iomap.type == IOMAP_HOLE)
1371 iomap_add_to_ioend(inode, file_offset, page, iop, wpc, wbc,
1376 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1377 WARN_ON_ONCE(!PageLocked(page));
1378 WARN_ON_ONCE(PageWriteback(page));
1379 WARN_ON_ONCE(PageDirty(page));
1382 * We cannot cancel the ioend directly here on error. We may have
1383 * already set other pages under writeback and hence we have to run I/O
1384 * completion to mark the error state of the pages under writeback
1387 if (unlikely(error)) {
1389 * Let the filesystem know what portion of the current page
1390 * failed to map. If the page wasn't been added to ioend, it
1391 * won't be affected by I/O completion and we must unlock it
1394 if (wpc->ops->discard_page)
1395 wpc->ops->discard_page(page, file_offset);
1397 ClearPageUptodate(page);
1403 set_page_writeback(page);
1407 * Preserve the original error if there was one, otherwise catch
1408 * submission errors here and propagate into subsequent ioend
1411 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1414 list_del_init(&ioend->io_list);
1415 error2 = iomap_submit_ioend(wpc, ioend, error);
1416 if (error2 && !error)
1421 * We can end up here with no error and nothing to write only if we race
1422 * with a partial page truncate on a sub-page block sized filesystem.
1425 end_page_writeback(page);
1427 mapping_set_error(page->mapping, error);
1432 * Write out a dirty page.
1434 * For delalloc space on the page we need to allocate space and flush it.
1435 * For unwritten space on the page we need to start the conversion to
1436 * regular allocated space.
1439 iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
1441 struct iomap_writepage_ctx *wpc = data;
1442 struct inode *inode = page->mapping->host;
1447 trace_iomap_writepage(inode, page_offset(page), PAGE_SIZE);
1450 * Refuse to write the page out if we are called from reclaim context.
1452 * This avoids stack overflows when called from deeply used stacks in
1453 * random callers for direct reclaim or memcg reclaim. We explicitly
1454 * allow reclaim from kswapd as the stack usage there is relatively low.
1456 * This should never happen except in the case of a VM regression so
1459 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1464 * Is this page beyond the end of the file?
1466 * The page index is less than the end_index, adjust the end_offset
1467 * to the highest offset that this page should represent.
1468 * -----------------------------------------------------
1469 * | file mapping | <EOF> |
1470 * -----------------------------------------------------
1471 * | Page ... | Page N-2 | Page N-1 | Page N | |
1472 * ^--------------------------------^----------|--------
1473 * | desired writeback range | see else |
1474 * ---------------------------------^------------------|
1476 offset = i_size_read(inode);
1477 end_index = offset >> PAGE_SHIFT;
1478 if (page->index < end_index)
1479 end_offset = (loff_t)(page->index + 1) << PAGE_SHIFT;
1482 * Check whether the page to write out is beyond or straddles
1484 * -------------------------------------------------------
1485 * | file mapping | <EOF> |
1486 * -------------------------------------------------------
1487 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1488 * ^--------------------------------^-----------|---------
1490 * ---------------------------------^-----------|--------|
1492 unsigned offset_into_page = offset & (PAGE_SIZE - 1);
1495 * Skip the page if it is fully outside i_size, e.g. due to a
1496 * truncate operation that is in progress. We must redirty the
1497 * page so that reclaim stops reclaiming it. Otherwise
1498 * iomap_vm_releasepage() is called on it and gets confused.
1500 * Note that the end_index is unsigned long, it would overflow
1501 * if the given offset is greater than 16TB on 32-bit system
1502 * and if we do check the page is fully outside i_size or not
1503 * via "if (page->index >= end_index + 1)" as "end_index + 1"
1504 * will be evaluated to 0. Hence this page will be redirtied
1505 * and be written out repeatedly which would result in an
1506 * infinite loop, the user program that perform this operation
1507 * will hang. Instead, we can verify this situation by checking
1508 * if the page to write is totally beyond the i_size or if it's
1509 * offset is just equal to the EOF.
1511 if (page->index > end_index ||
1512 (page->index == end_index && offset_into_page == 0))
1516 * The page straddles i_size. It must be zeroed out on each
1517 * and every writepage invocation because it may be mmapped.
1518 * "A file is mapped in multiples of the page size. For a file
1519 * that is not a multiple of the page size, the remaining
1520 * memory is zeroed when mapped, and writes to that region are
1521 * not written out to the file."
1523 zero_user_segment(page, offset_into_page, PAGE_SIZE);
1525 /* Adjust the end_offset to the end of file */
1526 end_offset = offset;
1529 return iomap_writepage_map(wpc, wbc, inode, page, end_offset);
1532 redirty_page_for_writepage(wbc, page);
1538 iomap_writepage(struct page *page, struct writeback_control *wbc,
1539 struct iomap_writepage_ctx *wpc,
1540 const struct iomap_writeback_ops *ops)
1545 ret = iomap_do_writepage(page, wbc, wpc);
1548 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1550 EXPORT_SYMBOL_GPL(iomap_writepage);
1553 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1554 struct iomap_writepage_ctx *wpc,
1555 const struct iomap_writeback_ops *ops)
1560 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1563 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1565 EXPORT_SYMBOL_GPL(iomap_writepages);
1567 static int __init iomap_init(void)
1569 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
1570 offsetof(struct iomap_ioend, io_inline_bio),
1573 fs_initcall(iomap_init);