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;
208 static loff_t iomap_read_inline_data(const struct iomap_iter *iter,
211 const struct iomap *iomap = iomap_iter_srcmap(iter);
212 size_t size = i_size_read(iter->inode) - iomap->offset;
213 size_t poff = offset_in_page(iomap->offset);
216 if (PageUptodate(page))
217 return PAGE_SIZE - poff;
219 if (WARN_ON_ONCE(size > PAGE_SIZE - poff))
221 if (WARN_ON_ONCE(size > PAGE_SIZE -
222 offset_in_page(iomap->inline_data)))
224 if (WARN_ON_ONCE(size > iomap->length))
227 iomap_page_create(iter->inode, page);
229 addr = kmap_local_page(page) + poff;
230 memcpy(addr, iomap->inline_data, size);
231 memset(addr + size, 0, PAGE_SIZE - poff - size);
233 iomap_set_range_uptodate(page, poff, PAGE_SIZE - poff);
234 return PAGE_SIZE - poff;
237 static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
240 const struct iomap *srcmap = iomap_iter_srcmap(iter);
242 return srcmap->type != IOMAP_MAPPED ||
243 (srcmap->flags & IOMAP_F_NEW) ||
244 pos >= i_size_read(iter->inode);
247 static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
248 struct iomap_readpage_ctx *ctx, loff_t offset)
250 const struct iomap *iomap = &iter->iomap;
251 loff_t pos = iter->pos + offset;
252 loff_t length = iomap_length(iter) - offset;
253 struct page *page = ctx->cur_page;
254 struct iomap_page *iop;
255 loff_t orig_pos = pos;
259 if (iomap->type == IOMAP_INLINE)
260 return min(iomap_read_inline_data(iter, page), length);
262 /* zero post-eof blocks as the page may be mapped */
263 iop = iomap_page_create(iter->inode, page);
264 iomap_adjust_read_range(iter->inode, iop, &pos, length, &poff, &plen);
268 if (iomap_block_needs_zeroing(iter, pos)) {
269 zero_user(page, poff, plen);
270 iomap_set_range_uptodate(page, poff, plen);
274 ctx->cur_page_in_bio = true;
276 atomic_add(plen, &iop->read_bytes_pending);
278 sector = iomap_sector(iomap, pos);
280 bio_end_sector(ctx->bio) != sector ||
281 bio_add_page(ctx->bio, page, plen, poff) != plen) {
282 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
283 gfp_t orig_gfp = gfp;
284 unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
287 submit_bio(ctx->bio);
289 if (ctx->rac) /* same as readahead_gfp_mask */
290 gfp |= __GFP_NORETRY | __GFP_NOWARN;
291 ctx->bio = bio_alloc(gfp, bio_max_segs(nr_vecs));
293 * If the bio_alloc fails, try it again for a single page to
294 * avoid having to deal with partial page reads. This emulates
295 * what do_mpage_readpage does.
298 ctx->bio = bio_alloc(orig_gfp, 1);
299 ctx->bio->bi_opf = REQ_OP_READ;
301 ctx->bio->bi_opf |= REQ_RAHEAD;
302 ctx->bio->bi_iter.bi_sector = sector;
303 bio_set_dev(ctx->bio, iomap->bdev);
304 ctx->bio->bi_end_io = iomap_read_end_io;
305 __bio_add_page(ctx->bio, page, plen, poff);
309 * Move the caller beyond our range so that it keeps making progress.
310 * For that, we have to include any leading non-uptodate ranges, but
311 * we can skip trailing ones as they will be handled in the next
314 return pos - orig_pos + plen;
318 iomap_readpage(struct page *page, const struct iomap_ops *ops)
320 struct iomap_iter iter = {
321 .inode = page->mapping->host,
322 .pos = page_offset(page),
325 struct iomap_readpage_ctx ctx = {
330 trace_iomap_readpage(page->mapping->host, 1);
332 while ((ret = iomap_iter(&iter, ops)) > 0)
333 iter.processed = iomap_readpage_iter(&iter, &ctx, 0);
340 WARN_ON_ONCE(!ctx.cur_page_in_bio);
342 WARN_ON_ONCE(ctx.cur_page_in_bio);
347 * Just like mpage_readahead and block_read_full_page, we always
348 * return 0 and just mark the page as PageError on errors. This
349 * should be cleaned up throughout the stack eventually.
353 EXPORT_SYMBOL_GPL(iomap_readpage);
355 static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
356 struct iomap_readpage_ctx *ctx)
358 loff_t length = iomap_length(iter);
361 for (done = 0; done < length; done += ret) {
362 if (ctx->cur_page && offset_in_page(iter->pos + done) == 0) {
363 if (!ctx->cur_page_in_bio)
364 unlock_page(ctx->cur_page);
365 put_page(ctx->cur_page);
366 ctx->cur_page = NULL;
368 if (!ctx->cur_page) {
369 ctx->cur_page = readahead_page(ctx->rac);
370 ctx->cur_page_in_bio = false;
372 ret = iomap_readpage_iter(iter, ctx, 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 iomap_iter iter = {
396 .inode = rac->mapping->host,
397 .pos = readahead_pos(rac),
398 .len = readahead_length(rac),
400 struct iomap_readpage_ctx ctx = {
404 trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
406 while (iomap_iter(&iter, ops) > 0)
407 iter.processed = iomap_readahead_iter(&iter, &ctx);
412 if (!ctx.cur_page_in_bio)
413 unlock_page(ctx.cur_page);
414 put_page(ctx.cur_page);
417 EXPORT_SYMBOL_GPL(iomap_readahead);
420 * iomap_is_partially_uptodate checks whether blocks within a page are
423 * Returns true if all blocks which correspond to a file portion
424 * we want to read within the page are uptodate.
427 iomap_is_partially_uptodate(struct page *page, unsigned long from,
430 struct iomap_page *iop = to_iomap_page(page);
431 struct inode *inode = page->mapping->host;
432 unsigned len, first, last;
435 /* Limit range to one page */
436 len = min_t(unsigned, PAGE_SIZE - from, count);
438 /* First and last blocks in range within page */
439 first = from >> inode->i_blkbits;
440 last = (from + len - 1) >> inode->i_blkbits;
443 for (i = first; i <= last; i++)
444 if (!test_bit(i, iop->uptodate))
451 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
454 iomap_releasepage(struct page *page, gfp_t gfp_mask)
456 trace_iomap_releasepage(page->mapping->host, page_offset(page),
460 * mm accommodates an old ext3 case where clean pages might not have had
461 * the dirty bit cleared. Thus, it can send actual dirty pages to
462 * ->releasepage() via shrink_active_list(); skip those here.
464 if (PageDirty(page) || PageWriteback(page))
466 iomap_page_release(page);
469 EXPORT_SYMBOL_GPL(iomap_releasepage);
472 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
474 trace_iomap_invalidatepage(page->mapping->host, offset, len);
477 * If we're invalidating the entire page, clear the dirty state from it
478 * and release it to avoid unnecessary buildup of the LRU.
480 if (offset == 0 && len == PAGE_SIZE) {
481 WARN_ON_ONCE(PageWriteback(page));
482 cancel_dirty_page(page);
483 iomap_page_release(page);
486 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
488 #ifdef CONFIG_MIGRATION
490 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
491 struct page *page, enum migrate_mode mode)
495 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
496 if (ret != MIGRATEPAGE_SUCCESS)
499 if (page_has_private(page))
500 attach_page_private(newpage, detach_page_private(page));
502 if (mode != MIGRATE_SYNC_NO_COPY)
503 migrate_page_copy(newpage, page);
505 migrate_page_states(newpage, page);
506 return MIGRATEPAGE_SUCCESS;
508 EXPORT_SYMBOL_GPL(iomap_migrate_page);
509 #endif /* CONFIG_MIGRATION */
512 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
514 loff_t i_size = i_size_read(inode);
517 * Only truncate newly allocated pages beyoned EOF, even if the
518 * write started inside the existing inode size.
520 if (pos + len > i_size)
521 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
525 iomap_read_page_sync(loff_t block_start, struct page *page, unsigned poff,
526 unsigned plen, const struct iomap *iomap)
531 bio_init(&bio, &bvec, 1);
532 bio.bi_opf = REQ_OP_READ;
533 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
534 bio_set_dev(&bio, iomap->bdev);
535 __bio_add_page(&bio, page, plen, poff);
536 return submit_bio_wait(&bio);
539 static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
540 unsigned len, struct page *page)
542 const struct iomap *srcmap = iomap_iter_srcmap(iter);
543 struct iomap_page *iop = iomap_page_create(iter->inode, page);
544 loff_t block_size = i_blocksize(iter->inode);
545 loff_t block_start = round_down(pos, block_size);
546 loff_t block_end = round_up(pos + len, block_size);
547 unsigned from = offset_in_page(pos), to = from + len, poff, plen;
549 if (PageUptodate(page))
551 ClearPageError(page);
554 iomap_adjust_read_range(iter->inode, iop, &block_start,
555 block_end - block_start, &poff, &plen);
559 if (!(iter->flags & IOMAP_UNSHARE) &&
560 (from <= poff || from >= poff + plen) &&
561 (to <= poff || to >= poff + plen))
564 if (iomap_block_needs_zeroing(iter, block_start)) {
565 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
567 zero_user_segments(page, poff, from, to, poff + plen);
569 int status = iomap_read_page_sync(block_start, page,
574 iomap_set_range_uptodate(page, poff, plen);
575 } while ((block_start += plen) < block_end);
580 static int iomap_write_begin_inline(const struct iomap_iter *iter,
585 /* needs more work for the tailpacking case; disable for now */
586 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
588 ret = iomap_read_inline_data(iter, page);
594 static int iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
595 unsigned len, struct page **pagep)
597 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
598 const struct iomap *srcmap = iomap_iter_srcmap(iter);
602 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
603 if (srcmap != &iter->iomap)
604 BUG_ON(pos + len > srcmap->offset + srcmap->length);
606 if (fatal_signal_pending(current))
609 if (page_ops && page_ops->page_prepare) {
610 status = page_ops->page_prepare(iter->inode, pos, len);
615 page = grab_cache_page_write_begin(iter->inode->i_mapping,
616 pos >> PAGE_SHIFT, AOP_FLAG_NOFS);
622 if (srcmap->type == IOMAP_INLINE)
623 status = iomap_write_begin_inline(iter, page);
624 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
625 status = __block_write_begin_int(page, pos, len, NULL, srcmap);
627 status = __iomap_write_begin(iter, pos, len, page);
629 if (unlikely(status))
638 iomap_write_failed(iter->inode, pos, len);
641 if (page_ops && page_ops->page_done)
642 page_ops->page_done(iter->inode, pos, 0, NULL);
646 static size_t __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
647 size_t copied, struct page *page)
649 flush_dcache_page(page);
652 * The blocks that were entirely written will now be uptodate, so we
653 * don't have to worry about a readpage reading them and overwriting a
654 * partial write. However, if we've encountered a short write and only
655 * partially written into a block, it will not be marked uptodate, so a
656 * readpage might come in and destroy our partial write.
658 * Do the simplest thing and just treat any short write to a
659 * non-uptodate page as a zero-length write, and force the caller to
660 * redo the whole thing.
662 if (unlikely(copied < len && !PageUptodate(page)))
664 iomap_set_range_uptodate(page, offset_in_page(pos), len);
665 __set_page_dirty_nobuffers(page);
669 static size_t iomap_write_end_inline(const struct iomap_iter *iter,
670 struct page *page, loff_t pos, size_t copied)
672 const struct iomap *iomap = &iter->iomap;
675 WARN_ON_ONCE(!PageUptodate(page));
676 BUG_ON(!iomap_inline_data_valid(iomap));
678 flush_dcache_page(page);
679 addr = kmap_local_page(page) + pos;
680 memcpy(iomap_inline_data(iomap, pos), addr, copied);
683 mark_inode_dirty(iter->inode);
687 /* Returns the number of bytes copied. May be 0. Cannot be an errno. */
688 static size_t iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
689 size_t copied, struct page *page)
691 const struct iomap_page_ops *page_ops = iter->iomap.page_ops;
692 const struct iomap *srcmap = iomap_iter_srcmap(iter);
693 loff_t old_size = iter->inode->i_size;
696 if (srcmap->type == IOMAP_INLINE) {
697 ret = iomap_write_end_inline(iter, page, pos, copied);
698 } else if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
699 ret = block_write_end(NULL, iter->inode->i_mapping, pos, len,
702 ret = __iomap_write_end(iter->inode, pos, len, copied, page);
706 * Update the in-memory inode size after copying the data into the page
707 * cache. It's up to the file system to write the updated size to disk,
708 * preferably after I/O completion so that no stale data is exposed.
710 if (pos + ret > old_size) {
711 i_size_write(iter->inode, pos + ret);
712 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
717 pagecache_isize_extended(iter->inode, old_size, pos);
718 if (page_ops && page_ops->page_done)
719 page_ops->page_done(iter->inode, pos, ret, page);
723 iomap_write_failed(iter->inode, pos, len);
727 static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
729 loff_t length = iomap_length(iter);
730 loff_t pos = iter->pos;
736 unsigned long offset; /* Offset into pagecache page */
737 unsigned long bytes; /* Bytes to write to page */
738 size_t copied; /* Bytes copied from user */
740 offset = offset_in_page(pos);
741 bytes = min_t(unsigned long, PAGE_SIZE - offset,
748 * Bring in the user page that we'll copy from _first_.
749 * Otherwise there's a nasty deadlock on copying from the
750 * same page as we're writing to, without it being marked
753 if (unlikely(fault_in_iov_iter_readable(i, bytes))) {
758 status = iomap_write_begin(iter, pos, bytes, &page);
759 if (unlikely(status))
762 if (mapping_writably_mapped(iter->inode->i_mapping))
763 flush_dcache_page(page);
765 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
767 status = iomap_write_end(iter, pos, bytes, copied, page);
769 if (unlikely(copied != status))
770 iov_iter_revert(i, copied - status);
773 if (unlikely(status == 0)) {
775 * A short copy made iomap_write_end() reject the
776 * thing entirely. Might be memory poisoning
777 * halfway through, might be a race with munmap,
778 * might be severe memory pressure.
788 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
789 } while (iov_iter_count(i) && length);
791 return written ? written : status;
795 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
796 const struct iomap_ops *ops)
798 struct iomap_iter iter = {
799 .inode = iocb->ki_filp->f_mapping->host,
801 .len = iov_iter_count(i),
802 .flags = IOMAP_WRITE,
806 while ((ret = iomap_iter(&iter, ops)) > 0)
807 iter.processed = iomap_write_iter(&iter, i);
808 if (iter.pos == iocb->ki_pos)
810 return iter.pos - iocb->ki_pos;
812 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
814 static loff_t iomap_unshare_iter(struct iomap_iter *iter)
816 struct iomap *iomap = &iter->iomap;
817 const struct iomap *srcmap = iomap_iter_srcmap(iter);
818 loff_t pos = iter->pos;
819 loff_t length = iomap_length(iter);
823 /* don't bother with blocks that are not shared to start with */
824 if (!(iomap->flags & IOMAP_F_SHARED))
826 /* don't bother with holes or unwritten extents */
827 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
831 unsigned long offset = offset_in_page(pos);
832 unsigned long bytes = min_t(loff_t, PAGE_SIZE - offset, length);
835 status = iomap_write_begin(iter, pos, bytes, &page);
836 if (unlikely(status))
839 status = iomap_write_end(iter, pos, bytes, bytes, page);
840 if (WARN_ON_ONCE(status == 0))
849 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
856 iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
857 const struct iomap_ops *ops)
859 struct iomap_iter iter = {
863 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
867 while ((ret = iomap_iter(&iter, ops)) > 0)
868 iter.processed = iomap_unshare_iter(&iter);
871 EXPORT_SYMBOL_GPL(iomap_file_unshare);
873 static s64 __iomap_zero_iter(struct iomap_iter *iter, loff_t pos, u64 length)
877 unsigned offset = offset_in_page(pos);
878 unsigned bytes = min_t(u64, PAGE_SIZE - offset, length);
880 status = iomap_write_begin(iter, pos, bytes, &page);
884 zero_user(page, offset, bytes);
885 mark_page_accessed(page);
887 return iomap_write_end(iter, pos, bytes, bytes, page);
890 static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
892 struct iomap *iomap = &iter->iomap;
893 const struct iomap *srcmap = iomap_iter_srcmap(iter);
894 loff_t pos = iter->pos;
895 loff_t length = iomap_length(iter);
898 /* already zeroed? we're done. */
899 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
905 if (IS_DAX(iter->inode))
906 bytes = dax_iomap_zero(pos, length, iomap);
908 bytes = __iomap_zero_iter(iter, pos, length);
917 } while (length > 0);
923 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
924 const struct iomap_ops *ops)
926 struct iomap_iter iter = {
934 while ((ret = iomap_iter(&iter, ops)) > 0)
935 iter.processed = iomap_zero_iter(&iter, did_zero);
938 EXPORT_SYMBOL_GPL(iomap_zero_range);
941 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
942 const struct iomap_ops *ops)
944 unsigned int blocksize = i_blocksize(inode);
945 unsigned int off = pos & (blocksize - 1);
947 /* Block boundary? Nothing to do */
950 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
952 EXPORT_SYMBOL_GPL(iomap_truncate_page);
954 static loff_t iomap_page_mkwrite_iter(struct iomap_iter *iter,
957 loff_t length = iomap_length(iter);
960 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
961 ret = __block_write_begin_int(page, iter->pos, length, NULL,
965 block_commit_write(page, 0, length);
967 WARN_ON_ONCE(!PageUptodate(page));
968 set_page_dirty(page);
974 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
976 struct iomap_iter iter = {
977 .inode = file_inode(vmf->vma->vm_file),
978 .flags = IOMAP_WRITE | IOMAP_FAULT,
980 struct page *page = vmf->page;
984 ret = page_mkwrite_check_truncate(page, iter.inode);
987 iter.pos = page_offset(page);
989 while ((ret = iomap_iter(&iter, ops)) > 0)
990 iter.processed = iomap_page_mkwrite_iter(&iter, page);
994 wait_for_stable_page(page);
995 return VM_FAULT_LOCKED;
998 return block_page_mkwrite_return(ret);
1000 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1003 iomap_finish_page_writeback(struct inode *inode, struct page *page,
1004 int error, unsigned int len)
1006 struct iomap_page *iop = to_iomap_page(page);
1010 mapping_set_error(inode->i_mapping, error);
1013 WARN_ON_ONCE(i_blocks_per_page(inode, page) > 1 && !iop);
1014 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) <= 0);
1016 if (!iop || atomic_sub_and_test(len, &iop->write_bytes_pending))
1017 end_page_writeback(page);
1021 * We're now finished for good with this ioend structure. Update the page
1022 * state, release holds on bios, and finally free up memory. Do not use the
1026 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1028 struct inode *inode = ioend->io_inode;
1029 struct bio *bio = &ioend->io_inline_bio;
1030 struct bio *last = ioend->io_bio, *next;
1031 u64 start = bio->bi_iter.bi_sector;
1032 loff_t offset = ioend->io_offset;
1033 bool quiet = bio_flagged(bio, BIO_QUIET);
1035 for (bio = &ioend->io_inline_bio; bio; bio = next) {
1037 struct bvec_iter_all iter_all;
1040 * For the last bio, bi_private points to the ioend, so we
1041 * need to explicitly end the iteration here.
1046 next = bio->bi_private;
1048 /* walk each page on bio, ending page IO on them */
1049 bio_for_each_segment_all(bv, bio, iter_all)
1050 iomap_finish_page_writeback(inode, bv->bv_page, error,
1054 /* The ioend has been freed by bio_put() */
1056 if (unlikely(error && !quiet)) {
1057 printk_ratelimited(KERN_ERR
1058 "%s: writeback error on inode %lu, offset %lld, sector %llu",
1059 inode->i_sb->s_id, inode->i_ino, offset, start);
1064 iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1066 struct list_head tmp;
1068 list_replace_init(&ioend->io_list, &tmp);
1069 iomap_finish_ioend(ioend, error);
1071 while (!list_empty(&tmp)) {
1072 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1073 list_del_init(&ioend->io_list);
1074 iomap_finish_ioend(ioend, error);
1077 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1080 * We can merge two adjacent ioends if they have the same set of work to do.
1083 iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1085 if (ioend->io_bio->bi_status != next->io_bio->bi_status)
1087 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1088 (next->io_flags & IOMAP_F_SHARED))
1090 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1091 (next->io_type == IOMAP_UNWRITTEN))
1093 if (ioend->io_offset + ioend->io_size != next->io_offset)
1099 iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1101 struct iomap_ioend *next;
1103 INIT_LIST_HEAD(&ioend->io_list);
1105 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1107 if (!iomap_ioend_can_merge(ioend, next))
1109 list_move_tail(&next->io_list, &ioend->io_list);
1110 ioend->io_size += next->io_size;
1113 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1116 iomap_ioend_compare(void *priv, const struct list_head *a,
1117 const struct list_head *b)
1119 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1120 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1122 if (ia->io_offset < ib->io_offset)
1124 if (ia->io_offset > ib->io_offset)
1130 iomap_sort_ioends(struct list_head *ioend_list)
1132 list_sort(NULL, ioend_list, iomap_ioend_compare);
1134 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1136 static void iomap_writepage_end_bio(struct bio *bio)
1138 struct iomap_ioend *ioend = bio->bi_private;
1140 iomap_finish_ioend(ioend, blk_status_to_errno(bio->bi_status));
1144 * Submit the final bio for an ioend.
1146 * If @error is non-zero, it means that we have a situation where some part of
1147 * the submission process has failed after we've marked pages for writeback
1148 * and unlocked them. In this situation, we need to fail the bio instead of
1149 * submitting it. This typically only happens on a filesystem shutdown.
1152 iomap_submit_ioend(struct iomap_writepage_ctx *wpc, struct iomap_ioend *ioend,
1155 ioend->io_bio->bi_private = ioend;
1156 ioend->io_bio->bi_end_io = iomap_writepage_end_bio;
1158 if (wpc->ops->prepare_ioend)
1159 error = wpc->ops->prepare_ioend(ioend, error);
1162 * If we're failing the IO now, just mark the ioend with an
1163 * error and finish it. This will run IO completion immediately
1164 * as there is only one reference to the ioend at this point in
1167 ioend->io_bio->bi_status = errno_to_blk_status(error);
1168 bio_endio(ioend->io_bio);
1172 submit_bio(ioend->io_bio);
1176 static struct iomap_ioend *
1177 iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
1178 loff_t offset, sector_t sector, struct writeback_control *wbc)
1180 struct iomap_ioend *ioend;
1183 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS, &iomap_ioend_bioset);
1184 bio_set_dev(bio, wpc->iomap.bdev);
1185 bio->bi_iter.bi_sector = sector;
1186 bio->bi_opf = REQ_OP_WRITE | wbc_to_write_flags(wbc);
1187 bio->bi_write_hint = inode->i_write_hint;
1188 wbc_init_bio(wbc, bio);
1190 ioend = container_of(bio, struct iomap_ioend, io_inline_bio);
1191 INIT_LIST_HEAD(&ioend->io_list);
1192 ioend->io_type = wpc->iomap.type;
1193 ioend->io_flags = wpc->iomap.flags;
1194 ioend->io_inode = inode;
1196 ioend->io_offset = offset;
1197 ioend->io_bio = bio;
1202 * Allocate a new bio, and chain the old bio to the new one.
1204 * Note that we have to perform the chaining in this unintuitive order
1205 * so that the bi_private linkage is set up in the right direction for the
1206 * traversal in iomap_finish_ioend().
1209 iomap_chain_bio(struct bio *prev)
1213 new = bio_alloc(GFP_NOFS, BIO_MAX_VECS);
1214 bio_copy_dev(new, prev);/* also copies over blkcg information */
1215 new->bi_iter.bi_sector = bio_end_sector(prev);
1216 new->bi_opf = prev->bi_opf;
1217 new->bi_write_hint = prev->bi_write_hint;
1219 bio_chain(prev, new);
1220 bio_get(prev); /* for iomap_finish_ioend */
1226 iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
1229 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1230 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1232 if (wpc->iomap.type != wpc->ioend->io_type)
1234 if (offset != wpc->ioend->io_offset + wpc->ioend->io_size)
1236 if (sector != bio_end_sector(wpc->ioend->io_bio))
1242 * Test to see if we have an existing ioend structure that we could append to
1243 * first; otherwise finish off the current ioend and start another.
1246 iomap_add_to_ioend(struct inode *inode, loff_t offset, struct page *page,
1247 struct iomap_page *iop, struct iomap_writepage_ctx *wpc,
1248 struct writeback_control *wbc, struct list_head *iolist)
1250 sector_t sector = iomap_sector(&wpc->iomap, offset);
1251 unsigned len = i_blocksize(inode);
1252 unsigned poff = offset & (PAGE_SIZE - 1);
1254 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, offset, sector)) {
1256 list_add(&wpc->ioend->io_list, iolist);
1257 wpc->ioend = iomap_alloc_ioend(inode, wpc, offset, sector, wbc);
1260 if (bio_add_page(wpc->ioend->io_bio, page, len, poff) != len) {
1261 wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
1262 __bio_add_page(wpc->ioend->io_bio, page, len, poff);
1266 atomic_add(len, &iop->write_bytes_pending);
1267 wpc->ioend->io_size += len;
1268 wbc_account_cgroup_owner(wbc, page, len);
1272 * We implement an immediate ioend submission policy here to avoid needing to
1273 * chain multiple ioends and hence nest mempool allocations which can violate
1274 * the forward progress guarantees we need to provide. The current ioend we're
1275 * adding blocks to is cached in the writepage context, and if the new block
1276 * doesn't append to the cached ioend, it will create a new ioend and cache that
1279 * If a new ioend is created and cached, the old ioend is returned and queued
1280 * locally for submission once the entire page is processed or an error has been
1281 * detected. While ioends are submitted immediately after they are completed,
1282 * batching optimisations are provided by higher level block plugging.
1284 * At the end of a writeback pass, there will be a cached ioend remaining on the
1285 * writepage context that the caller will need to submit.
1288 iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1289 struct writeback_control *wbc, struct inode *inode,
1290 struct page *page, u64 end_offset)
1292 struct iomap_page *iop = iomap_page_create(inode, page);
1293 struct iomap_ioend *ioend, *next;
1294 unsigned len = i_blocksize(inode);
1295 u64 file_offset; /* file offset of page */
1296 int error = 0, count = 0, i;
1297 LIST_HEAD(submit_list);
1299 WARN_ON_ONCE(iop && atomic_read(&iop->write_bytes_pending) != 0);
1302 * Walk through the page to find areas to write back. If we run off the
1303 * end of the current map or find the current map invalid, grab a new
1306 for (i = 0, file_offset = page_offset(page);
1307 i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
1308 i++, file_offset += len) {
1309 if (iop && !test_bit(i, iop->uptodate))
1312 error = wpc->ops->map_blocks(wpc, inode, file_offset);
1315 if (WARN_ON_ONCE(wpc->iomap.type == IOMAP_INLINE))
1317 if (wpc->iomap.type == IOMAP_HOLE)
1319 iomap_add_to_ioend(inode, file_offset, page, iop, wpc, wbc,
1324 WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
1325 WARN_ON_ONCE(!PageLocked(page));
1326 WARN_ON_ONCE(PageWriteback(page));
1327 WARN_ON_ONCE(PageDirty(page));
1330 * We cannot cancel the ioend directly here on error. We may have
1331 * already set other pages under writeback and hence we have to run I/O
1332 * completion to mark the error state of the pages under writeback
1335 if (unlikely(error)) {
1337 * Let the filesystem know what portion of the current page
1338 * failed to map. If the page hasn't been added to ioend, it
1339 * won't be affected by I/O completion and we must unlock it
1342 if (wpc->ops->discard_page)
1343 wpc->ops->discard_page(page, file_offset);
1345 ClearPageUptodate(page);
1351 set_page_writeback(page);
1355 * Preserve the original error if there was one; catch
1356 * submission errors here and propagate into subsequent ioend
1359 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
1362 list_del_init(&ioend->io_list);
1363 error2 = iomap_submit_ioend(wpc, ioend, error);
1364 if (error2 && !error)
1369 * We can end up here with no error and nothing to write only if we race
1370 * with a partial page truncate on a sub-page block sized filesystem.
1373 end_page_writeback(page);
1375 mapping_set_error(page->mapping, error);
1380 * Write out a dirty page.
1382 * For delalloc space on the page, we need to allocate space and flush it.
1383 * For unwritten space on the page, we need to start the conversion to
1384 * regular allocated space.
1387 iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
1389 struct iomap_writepage_ctx *wpc = data;
1390 struct inode *inode = page->mapping->host;
1395 trace_iomap_writepage(inode, page_offset(page), PAGE_SIZE);
1398 * Refuse to write the page out if we're called from reclaim context.
1400 * This avoids stack overflows when called from deeply used stacks in
1401 * random callers for direct reclaim or memcg reclaim. We explicitly
1402 * allow reclaim from kswapd as the stack usage there is relatively low.
1404 * This should never happen except in the case of a VM regression so
1407 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1412 * Is this page beyond the end of the file?
1414 * The page index is less than the end_index, adjust the end_offset
1415 * to the highest offset that this page should represent.
1416 * -----------------------------------------------------
1417 * | file mapping | <EOF> |
1418 * -----------------------------------------------------
1419 * | Page ... | Page N-2 | Page N-1 | Page N | |
1420 * ^--------------------------------^----------|--------
1421 * | desired writeback range | see else |
1422 * ---------------------------------^------------------|
1424 offset = i_size_read(inode);
1425 end_index = offset >> PAGE_SHIFT;
1426 if (page->index < end_index)
1427 end_offset = (loff_t)(page->index + 1) << PAGE_SHIFT;
1430 * Check whether the page to write out is beyond or straddles
1432 * -------------------------------------------------------
1433 * | file mapping | <EOF> |
1434 * -------------------------------------------------------
1435 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1436 * ^--------------------------------^-----------|---------
1438 * ---------------------------------^-----------|--------|
1440 unsigned offset_into_page = offset & (PAGE_SIZE - 1);
1443 * Skip the page if it's fully outside i_size, e.g. due to a
1444 * truncate operation that's in progress. We must redirty the
1445 * page so that reclaim stops reclaiming it. Otherwise
1446 * iomap_vm_releasepage() is called on it and gets confused.
1448 * Note that the end_index is unsigned long. If the given
1449 * offset is greater than 16TB on a 32-bit system then if we
1450 * checked if the page is fully outside i_size with
1451 * "if (page->index >= end_index + 1)", "end_index + 1" would
1452 * overflow and evaluate to 0. Hence this page would be
1453 * redirtied and written out repeatedly, which would result in
1454 * an infinite loop; the user program performing this operation
1455 * would hang. Instead, we can detect this situation by
1456 * checking if the page is totally beyond i_size or if its
1457 * offset is just equal to the EOF.
1459 if (page->index > end_index ||
1460 (page->index == end_index && offset_into_page == 0))
1464 * The page straddles i_size. It must be zeroed out on each
1465 * and every writepage invocation because it may be mmapped.
1466 * "A file is mapped in multiples of the page size. For a file
1467 * that is not a multiple of the page size, the remaining
1468 * memory is zeroed when mapped, and writes to that region are
1469 * not written out to the file."
1471 zero_user_segment(page, offset_into_page, PAGE_SIZE);
1473 /* Adjust the end_offset to the end of file */
1474 end_offset = offset;
1477 return iomap_writepage_map(wpc, wbc, inode, page, end_offset);
1480 redirty_page_for_writepage(wbc, page);
1486 iomap_writepage(struct page *page, struct writeback_control *wbc,
1487 struct iomap_writepage_ctx *wpc,
1488 const struct iomap_writeback_ops *ops)
1493 ret = iomap_do_writepage(page, wbc, wpc);
1496 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1498 EXPORT_SYMBOL_GPL(iomap_writepage);
1501 iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
1502 struct iomap_writepage_ctx *wpc,
1503 const struct iomap_writeback_ops *ops)
1508 ret = write_cache_pages(mapping, wbc, iomap_do_writepage, wpc);
1511 return iomap_submit_ioend(wpc, wpc->ioend, ret);
1513 EXPORT_SYMBOL_GPL(iomap_writepages);
1515 static int __init iomap_init(void)
1517 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
1518 offsetof(struct iomap_ioend, io_inline_bio),
1521 fs_initcall(iomap_init);