2 * Copyright (C) 2010 Red Hat, Inc.
3 * Copyright (c) 2016-2018 Christoph Hellwig.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #include <linux/module.h>
15 #include <linux/compiler.h>
17 #include <linux/iomap.h>
18 #include <linux/uaccess.h>
19 #include <linux/gfp.h>
20 #include <linux/migrate.h>
22 #include <linux/mm_inline.h>
23 #include <linux/swap.h>
24 #include <linux/pagemap.h>
25 #include <linux/pagevec.h>
26 #include <linux/file.h>
27 #include <linux/uio.h>
28 #include <linux/backing-dev.h>
29 #include <linux/buffer_head.h>
30 #include <linux/task_io_accounting_ops.h>
31 #include <linux/dax.h>
32 #include <linux/sched/signal.h>
33 #include <linux/swap.h>
38 * Execute a iomap write on a segment of the mapping that spans a
39 * contiguous range of pages that have identical block mapping state.
41 * This avoids the need to map pages individually, do individual allocations
42 * for each page and most importantly avoid the need for filesystem specific
43 * locking per page. Instead, all the operations are amortised over the entire
44 * range of pages. It is assumed that the filesystems will lock whatever
45 * resources they require in the iomap_begin call, and release them in the
49 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
50 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
52 struct iomap iomap = { 0 };
53 loff_t written = 0, ret;
56 * Need to map a range from start position for length bytes. This can
57 * span multiple pages - it is only guaranteed to return a range of a
58 * single type of pages (e.g. all into a hole, all mapped or all
59 * unwritten). Failure at this point has nothing to undo.
61 * If allocation is required for this range, reserve the space now so
62 * that the allocation is guaranteed to succeed later on. Once we copy
63 * the data into the page cache pages, then we cannot fail otherwise we
64 * expose transient stale data. If the reserve fails, we can safely
65 * back out at this point as there is nothing to undo.
67 ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
70 if (WARN_ON(iomap.offset > pos))
72 if (WARN_ON(iomap.length == 0))
76 * Cut down the length to the one actually provided by the filesystem,
77 * as it might not be able to give us the whole size that we requested.
79 if (iomap.offset + iomap.length < pos + length)
80 length = iomap.offset + iomap.length - pos;
83 * Now that we have guaranteed that the space allocation will succeed.
84 * we can do the copy-in page by page without having to worry about
85 * failures exposing transient data.
87 written = actor(inode, pos, length, data, &iomap);
90 * Now the data has been copied, commit the range we've copied. This
91 * should not fail unless the filesystem has had a fatal error.
94 ret = ops->iomap_end(inode, pos, length,
95 written > 0 ? written : 0,
99 return written ? written : ret;
103 iomap_sector(struct iomap *iomap, loff_t pos)
105 return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
108 static struct iomap_page *
109 iomap_page_create(struct inode *inode, struct page *page)
111 struct iomap_page *iop = to_iomap_page(page);
112 unsigned int nr_blocks = PAGE_SIZE / i_blocksize(inode);
114 if (iop || i_blocksize(inode) == PAGE_SIZE)
117 iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
118 atomic_set(&iop->read_count, 0);
119 atomic_set(&iop->write_count, 0);
120 spin_lock_init(&iop->uptodate_lock);
121 bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
122 if (PageUptodate(page))
123 bitmap_fill(iop->uptodate, nr_blocks);
126 * migrate_page_move_mapping() assumes that pages with private data have
127 * their count elevated by 1.
130 set_page_private(page, (unsigned long)iop);
131 SetPagePrivate(page);
136 iomap_page_release(struct page *page)
138 struct iomap_page *iop = to_iomap_page(page);
142 WARN_ON_ONCE(atomic_read(&iop->read_count));
143 WARN_ON_ONCE(atomic_read(&iop->write_count));
144 ClearPagePrivate(page);
145 set_page_private(page, 0);
151 * Calculate the range inside the page that we actually need to read.
154 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
155 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
157 loff_t orig_pos = *pos;
158 loff_t isize = i_size_read(inode);
159 unsigned block_bits = inode->i_blkbits;
160 unsigned block_size = (1 << block_bits);
161 unsigned poff = offset_in_page(*pos);
162 unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
163 unsigned first = poff >> block_bits;
164 unsigned last = (poff + plen - 1) >> block_bits;
167 * If the block size is smaller than the page size we need to check the
168 * per-block uptodate status and adjust the offset and length if needed
169 * to avoid reading in already uptodate ranges.
174 /* move forward for each leading block marked uptodate */
175 for (i = first; i <= last; i++) {
176 if (!test_bit(i, iop->uptodate))
184 /* truncate len if we find any trailing uptodate block(s) */
185 for ( ; i <= last; i++) {
186 if (test_bit(i, iop->uptodate)) {
187 plen -= (last - i + 1) * block_size;
195 * If the extent spans the block that contains the i_size we need to
196 * handle both halves separately so that we properly zero data in the
197 * page cache for blocks that are entirely outside of i_size.
199 if (orig_pos <= isize && orig_pos + length > isize) {
200 unsigned end = offset_in_page(isize - 1) >> block_bits;
202 if (first <= end && last > end)
203 plen -= (last - end) * block_size;
211 iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
213 struct iomap_page *iop = to_iomap_page(page);
214 struct inode *inode = page->mapping->host;
215 unsigned first = off >> inode->i_blkbits;
216 unsigned last = (off + len - 1) >> inode->i_blkbits;
217 bool uptodate = true;
221 spin_lock_irqsave(&iop->uptodate_lock, flags);
222 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
223 if (i >= first && i <= last)
224 set_bit(i, iop->uptodate);
225 else if (!test_bit(i, iop->uptodate))
230 SetPageUptodate(page);
231 spin_unlock_irqrestore(&iop->uptodate_lock, flags);
235 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
240 if (page_has_private(page))
241 iomap_iop_set_range_uptodate(page, off, len);
243 SetPageUptodate(page);
247 iomap_read_finish(struct iomap_page *iop, struct page *page)
249 if (!iop || atomic_dec_and_test(&iop->read_count))
254 iomap_read_page_end_io(struct bio_vec *bvec, int error)
256 struct page *page = bvec->bv_page;
257 struct iomap_page *iop = to_iomap_page(page);
259 if (unlikely(error)) {
260 ClearPageUptodate(page);
263 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
266 iomap_read_finish(iop, page);
270 iomap_read_inline_data(struct inode *inode, struct page *page,
273 size_t size = i_size_read(inode);
276 if (PageUptodate(page))
280 BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
282 addr = kmap_atomic(page);
283 memcpy(addr, iomap->inline_data, size);
284 memset(addr + size, 0, PAGE_SIZE - size);
286 SetPageUptodate(page);
290 iomap_read_end_io(struct bio *bio)
292 int error = blk_status_to_errno(bio->bi_status);
293 struct bio_vec *bvec;
296 bio_for_each_segment_all(bvec, bio, i)
297 iomap_read_page_end_io(bvec, error);
301 struct iomap_readpage_ctx {
302 struct page *cur_page;
303 bool cur_page_in_bio;
306 struct list_head *pages;
310 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
313 struct iomap_readpage_ctx *ctx = data;
314 struct page *page = ctx->cur_page;
315 struct iomap_page *iop = iomap_page_create(inode, page);
316 bool is_contig = false;
317 loff_t orig_pos = pos;
321 if (iomap->type == IOMAP_INLINE) {
323 iomap_read_inline_data(inode, page, iomap);
327 /* zero post-eof blocks as the page may be mapped */
328 iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
332 if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
333 zero_user(page, poff, plen);
334 iomap_set_range_uptodate(page, poff, plen);
338 ctx->cur_page_in_bio = true;
341 * Try to merge into a previous segment if we can.
343 sector = iomap_sector(iomap, pos);
344 if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
345 if (__bio_try_merge_page(ctx->bio, page, plen, poff))
351 * If we start a new segment we need to increase the read count, and we
352 * need to do so before submitting any previous full bio to make sure
353 * that we don't prematurely unlock the page.
356 atomic_inc(&iop->read_count);
358 if (!ctx->bio || !is_contig || bio_full(ctx->bio)) {
359 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
360 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
363 submit_bio(ctx->bio);
365 if (ctx->is_readahead) /* same as readahead_gfp_mask */
366 gfp |= __GFP_NORETRY | __GFP_NOWARN;
367 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
368 ctx->bio->bi_opf = REQ_OP_READ;
369 if (ctx->is_readahead)
370 ctx->bio->bi_opf |= REQ_RAHEAD;
371 ctx->bio->bi_iter.bi_sector = sector;
372 bio_set_dev(ctx->bio, iomap->bdev);
373 ctx->bio->bi_end_io = iomap_read_end_io;
376 __bio_add_page(ctx->bio, page, plen, poff);
379 * Move the caller beyond our range so that it keeps making progress.
380 * For that we have to include any leading non-uptodate ranges, but
381 * we can skip trailing ones as they will be handled in the next
384 return pos - orig_pos + plen;
388 iomap_readpage(struct page *page, const struct iomap_ops *ops)
390 struct iomap_readpage_ctx ctx = { .cur_page = page };
391 struct inode *inode = page->mapping->host;
395 for (poff = 0; poff < PAGE_SIZE; poff += ret) {
396 ret = iomap_apply(inode, page_offset(page) + poff,
397 PAGE_SIZE - poff, 0, ops, &ctx,
398 iomap_readpage_actor);
400 WARN_ON_ONCE(ret == 0);
408 WARN_ON_ONCE(!ctx.cur_page_in_bio);
410 WARN_ON_ONCE(ctx.cur_page_in_bio);
415 * Just like mpage_readpages and block_read_full_page we always
416 * return 0 and just mark the page as PageError on errors. This
417 * should be cleaned up all through the stack eventually.
421 EXPORT_SYMBOL_GPL(iomap_readpage);
424 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
425 loff_t length, loff_t *done)
427 while (!list_empty(pages)) {
428 struct page *page = lru_to_page(pages);
430 if (page_offset(page) >= (u64)pos + length)
433 list_del(&page->lru);
434 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
439 * If we already have a page in the page cache at index we are
440 * done. Upper layers don't care if it is uptodate after the
441 * readpages call itself as every page gets checked again once
452 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
453 void *data, struct iomap *iomap)
455 struct iomap_readpage_ctx *ctx = data;
458 for (done = 0; done < length; done += ret) {
459 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
460 if (!ctx->cur_page_in_bio)
461 unlock_page(ctx->cur_page);
462 put_page(ctx->cur_page);
463 ctx->cur_page = NULL;
465 if (!ctx->cur_page) {
466 ctx->cur_page = iomap_next_page(inode, ctx->pages,
470 ctx->cur_page_in_bio = false;
472 ret = iomap_readpage_actor(inode, pos + done, length - done,
480 iomap_readpages(struct address_space *mapping, struct list_head *pages,
481 unsigned nr_pages, const struct iomap_ops *ops)
483 struct iomap_readpage_ctx ctx = {
485 .is_readahead = true,
487 loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
488 loff_t last = page_offset(list_entry(pages->next, struct page, lru));
489 loff_t length = last - pos + PAGE_SIZE, ret = 0;
492 ret = iomap_apply(mapping->host, pos, length, 0, ops,
493 &ctx, iomap_readpages_actor);
495 WARN_ON_ONCE(ret == 0);
506 if (!ctx.cur_page_in_bio)
507 unlock_page(ctx.cur_page);
508 put_page(ctx.cur_page);
512 * Check that we didn't lose a page due to the arcance calling
515 WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
518 EXPORT_SYMBOL_GPL(iomap_readpages);
521 * iomap_is_partially_uptodate checks whether blocks within a page are
524 * Returns true if all blocks which correspond to a file portion
525 * we want to read within the page are uptodate.
528 iomap_is_partially_uptodate(struct page *page, unsigned long from,
531 struct iomap_page *iop = to_iomap_page(page);
532 struct inode *inode = page->mapping->host;
533 unsigned len, first, last;
536 /* Limit range to one page */
537 len = min_t(unsigned, PAGE_SIZE - from, count);
539 /* First and last blocks in range within page */
540 first = from >> inode->i_blkbits;
541 last = (from + len - 1) >> inode->i_blkbits;
544 for (i = first; i <= last; i++)
545 if (!test_bit(i, iop->uptodate))
552 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
555 iomap_releasepage(struct page *page, gfp_t gfp_mask)
558 * mm accommodates an old ext3 case where clean pages might not have had
559 * the dirty bit cleared. Thus, it can send actual dirty pages to
560 * ->releasepage() via shrink_active_list(), skip those here.
562 if (PageDirty(page) || PageWriteback(page))
564 iomap_page_release(page);
567 EXPORT_SYMBOL_GPL(iomap_releasepage);
570 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
573 * If we are invalidating the entire page, clear the dirty state from it
574 * and release it to avoid unnecessary buildup of the LRU.
576 if (offset == 0 && len == PAGE_SIZE) {
577 WARN_ON_ONCE(PageWriteback(page));
578 cancel_dirty_page(page);
579 iomap_page_release(page);
582 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
584 #ifdef CONFIG_MIGRATION
586 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
587 struct page *page, enum migrate_mode mode)
591 ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
592 if (ret != MIGRATEPAGE_SUCCESS)
595 if (page_has_private(page)) {
596 ClearPagePrivate(page);
598 set_page_private(newpage, page_private(page));
599 set_page_private(page, 0);
601 SetPagePrivate(newpage);
604 if (mode != MIGRATE_SYNC_NO_COPY)
605 migrate_page_copy(newpage, page);
607 migrate_page_states(newpage, page);
608 return MIGRATEPAGE_SUCCESS;
610 EXPORT_SYMBOL_GPL(iomap_migrate_page);
611 #endif /* CONFIG_MIGRATION */
614 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
616 loff_t i_size = i_size_read(inode);
619 * Only truncate newly allocated pages beyoned EOF, even if the
620 * write started inside the existing inode size.
622 if (pos + len > i_size)
623 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
627 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
628 unsigned poff, unsigned plen, unsigned from, unsigned to,
634 if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
635 zero_user_segments(page, poff, from, to, poff + plen);
636 iomap_set_range_uptodate(page, poff, plen);
640 bio_init(&bio, &bvec, 1);
641 bio.bi_opf = REQ_OP_READ;
642 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
643 bio_set_dev(&bio, iomap->bdev);
644 __bio_add_page(&bio, page, plen, poff);
645 return submit_bio_wait(&bio);
649 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
650 struct page *page, struct iomap *iomap)
652 struct iomap_page *iop = iomap_page_create(inode, page);
653 loff_t block_size = i_blocksize(inode);
654 loff_t block_start = pos & ~(block_size - 1);
655 loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
656 unsigned from = offset_in_page(pos), to = from + len, poff, plen;
659 if (PageUptodate(page))
663 iomap_adjust_read_range(inode, iop, &block_start,
664 block_end - block_start, &poff, &plen);
668 if ((from > poff && from < poff + plen) ||
669 (to > poff && to < poff + plen)) {
670 status = iomap_read_page_sync(inode, block_start, page,
671 poff, plen, from, to, iomap);
676 } while ((block_start += plen) < block_end);
682 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
683 struct page **pagep, struct iomap *iomap)
685 pgoff_t index = pos >> PAGE_SHIFT;
689 BUG_ON(pos + len > iomap->offset + iomap->length);
691 if (fatal_signal_pending(current))
694 page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
698 if (iomap->type == IOMAP_INLINE)
699 iomap_read_inline_data(inode, page, iomap);
700 else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
701 status = __block_write_begin_int(page, pos, len, NULL, iomap);
703 status = __iomap_write_begin(inode, pos, len, page, iomap);
704 if (unlikely(status)) {
709 iomap_write_failed(inode, pos, len);
717 iomap_set_page_dirty(struct page *page)
719 struct address_space *mapping = page_mapping(page);
722 if (unlikely(!mapping))
723 return !TestSetPageDirty(page);
726 * Lock out page->mem_cgroup migration to keep PageDirty
727 * synchronized with per-memcg dirty page counters.
729 lock_page_memcg(page);
730 newly_dirty = !TestSetPageDirty(page);
732 __set_page_dirty(page, mapping, 0);
733 unlock_page_memcg(page);
736 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
739 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
742 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
743 unsigned copied, struct page *page, struct iomap *iomap)
745 flush_dcache_page(page);
748 * The blocks that were entirely written will now be uptodate, so we
749 * don't have to worry about a readpage reading them and overwriting a
750 * partial write. However if we have encountered a short write and only
751 * partially written into a block, it will not be marked uptodate, so a
752 * readpage might come in and destroy our partial write.
754 * Do the simplest thing, and just treat any short write to a non
755 * uptodate page as a zero-length write, and force the caller to redo
758 if (unlikely(copied < len && !PageUptodate(page))) {
761 iomap_set_range_uptodate(page, offset_in_page(pos), len);
762 iomap_set_page_dirty(page);
764 return __generic_write_end(inode, pos, copied, page);
768 iomap_write_end_inline(struct inode *inode, struct page *page,
769 struct iomap *iomap, loff_t pos, unsigned copied)
773 WARN_ON_ONCE(!PageUptodate(page));
774 BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
776 addr = kmap_atomic(page);
777 memcpy(iomap->inline_data + pos, addr + pos, copied);
780 mark_inode_dirty(inode);
781 __generic_write_end(inode, pos, copied, page);
786 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
787 unsigned copied, struct page *page, struct iomap *iomap)
791 if (iomap->type == IOMAP_INLINE) {
792 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
793 } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
794 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
797 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
800 if (iomap->page_done)
801 iomap->page_done(inode, pos, copied, page, iomap);
804 iomap_write_failed(inode, pos, len);
809 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
812 struct iov_iter *i = data;
815 unsigned int flags = AOP_FLAG_NOFS;
819 unsigned long offset; /* Offset into pagecache page */
820 unsigned long bytes; /* Bytes to write to page */
821 size_t copied; /* Bytes copied from user */
823 offset = offset_in_page(pos);
824 bytes = min_t(unsigned long, PAGE_SIZE - offset,
831 * Bring in the user page that we will copy from _first_.
832 * Otherwise there's a nasty deadlock on copying from the
833 * same page as we're writing to, without it being marked
836 * Not only is this an optimisation, but it is also required
837 * to check that the address is actually valid, when atomic
838 * usercopies are used, below.
840 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
845 status = iomap_write_begin(inode, pos, bytes, flags, &page,
847 if (unlikely(status))
850 if (mapping_writably_mapped(inode->i_mapping))
851 flush_dcache_page(page);
853 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
855 flush_dcache_page(page);
857 status = iomap_write_end(inode, pos, bytes, copied, page,
859 if (unlikely(status < 0))
865 iov_iter_advance(i, copied);
866 if (unlikely(copied == 0)) {
868 * If we were unable to copy any data at all, we must
869 * fall back to a single segment length write.
871 * If we didn't fallback here, we could livelock
872 * because not all segments in the iov can be copied at
873 * once without a pagefault.
875 bytes = min_t(unsigned long, PAGE_SIZE - offset,
876 iov_iter_single_seg_count(i));
883 balance_dirty_pages_ratelimited(inode->i_mapping);
884 } while (iov_iter_count(i) && length);
886 return written ? written : status;
890 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
891 const struct iomap_ops *ops)
893 struct inode *inode = iocb->ki_filp->f_mapping->host;
894 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
896 while (iov_iter_count(iter)) {
897 ret = iomap_apply(inode, pos, iov_iter_count(iter),
898 IOMAP_WRITE, ops, iter, iomap_write_actor);
905 return written ? written : ret;
907 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
910 __iomap_read_page(struct inode *inode, loff_t offset)
912 struct address_space *mapping = inode->i_mapping;
915 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
918 if (!PageUptodate(page)) {
920 return ERR_PTR(-EIO);
926 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
933 struct page *page, *rpage;
934 unsigned long offset; /* Offset into pagecache page */
935 unsigned long bytes; /* Bytes to write to page */
937 offset = offset_in_page(pos);
938 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
940 rpage = __iomap_read_page(inode, pos);
942 return PTR_ERR(rpage);
944 status = iomap_write_begin(inode, pos, bytes,
945 AOP_FLAG_NOFS, &page, iomap);
947 if (unlikely(status))
950 WARN_ON_ONCE(!PageUptodate(page));
952 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
953 if (unlikely(status <= 0)) {
954 if (WARN_ON_ONCE(status == 0))
965 balance_dirty_pages_ratelimited(inode->i_mapping);
972 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
973 const struct iomap_ops *ops)
978 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
988 EXPORT_SYMBOL_GPL(iomap_file_dirty);
990 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
991 unsigned bytes, struct iomap *iomap)
996 status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
1001 zero_user(page, offset, bytes);
1002 mark_page_accessed(page);
1004 return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
1007 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
1008 struct iomap *iomap)
1010 return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
1011 iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
1015 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
1016 void *data, struct iomap *iomap)
1018 bool *did_zero = data;
1022 /* already zeroed? we're done. */
1023 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1027 unsigned offset, bytes;
1029 offset = offset_in_page(pos);
1030 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1033 status = iomap_dax_zero(pos, offset, bytes, iomap);
1035 status = iomap_zero(inode, pos, offset, bytes, iomap);
1044 } while (count > 0);
1050 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1051 const struct iomap_ops *ops)
1056 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1057 ops, did_zero, iomap_zero_range_actor);
1067 EXPORT_SYMBOL_GPL(iomap_zero_range);
1070 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1071 const struct iomap_ops *ops)
1073 unsigned int blocksize = i_blocksize(inode);
1074 unsigned int off = pos & (blocksize - 1);
1076 /* Block boundary? Nothing to do */
1079 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1081 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1084 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1085 void *data, struct iomap *iomap)
1087 struct page *page = data;
1090 if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1091 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1094 block_commit_write(page, 0, length);
1096 WARN_ON_ONCE(!PageUptodate(page));
1097 iomap_page_create(inode, page);
1098 set_page_dirty(page);
1104 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1106 struct page *page = vmf->page;
1107 struct inode *inode = file_inode(vmf->vma->vm_file);
1108 unsigned long length;
1109 loff_t offset, size;
1113 size = i_size_read(inode);
1114 if ((page->mapping != inode->i_mapping) ||
1115 (page_offset(page) > size)) {
1116 /* We overload EFAULT to mean page got truncated */
1121 /* page is wholly or partially inside EOF */
1122 if (((page->index + 1) << PAGE_SHIFT) > size)
1123 length = offset_in_page(size);
1127 offset = page_offset(page);
1128 while (length > 0) {
1129 ret = iomap_apply(inode, offset, length,
1130 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1131 iomap_page_mkwrite_actor);
1132 if (unlikely(ret <= 0))
1138 wait_for_stable_page(page);
1139 return VM_FAULT_LOCKED;
1142 return block_page_mkwrite_return(ret);
1144 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1147 struct fiemap_extent_info *fi;
1151 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1152 struct iomap *iomap, u32 flags)
1154 switch (iomap->type) {
1158 case IOMAP_DELALLOC:
1159 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1163 case IOMAP_UNWRITTEN:
1164 flags |= FIEMAP_EXTENT_UNWRITTEN;
1167 flags |= FIEMAP_EXTENT_DATA_INLINE;
1171 if (iomap->flags & IOMAP_F_MERGED)
1172 flags |= FIEMAP_EXTENT_MERGED;
1173 if (iomap->flags & IOMAP_F_SHARED)
1174 flags |= FIEMAP_EXTENT_SHARED;
1176 return fiemap_fill_next_extent(fi, iomap->offset,
1177 iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1178 iomap->length, flags);
1182 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1183 struct iomap *iomap)
1185 struct fiemap_ctx *ctx = data;
1186 loff_t ret = length;
1188 if (iomap->type == IOMAP_HOLE)
1191 ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1194 case 0: /* success */
1196 case 1: /* extent array full */
1203 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1204 loff_t start, loff_t len, const struct iomap_ops *ops)
1206 struct fiemap_ctx ctx;
1209 memset(&ctx, 0, sizeof(ctx));
1211 ctx.prev.type = IOMAP_HOLE;
1213 ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1217 if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1218 ret = filemap_write_and_wait(inode->i_mapping);
1224 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1225 iomap_fiemap_actor);
1226 /* inode with no (attribute) mapping will give ENOENT */
1238 if (ctx.prev.type != IOMAP_HOLE) {
1239 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1246 EXPORT_SYMBOL_GPL(iomap_fiemap);
1249 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1250 * Returns true if found and updates @lastoff to the offset in file.
1253 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1256 const struct address_space_operations *ops = inode->i_mapping->a_ops;
1257 unsigned int bsize = i_blocksize(inode), off;
1258 bool seek_data = whence == SEEK_DATA;
1259 loff_t poff = page_offset(page);
1261 if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1264 if (*lastoff < poff) {
1266 * Last offset smaller than the start of the page means we found
1269 if (whence == SEEK_HOLE)
1275 * Just check the page unless we can and should check block ranges:
1277 if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1278 return PageUptodate(page) == seek_data;
1281 if (unlikely(page->mapping != inode->i_mapping))
1282 goto out_unlock_not_found;
1284 for (off = 0; off < PAGE_SIZE; off += bsize) {
1285 if (offset_in_page(*lastoff) >= off + bsize)
1287 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1291 *lastoff = poff + off + bsize;
1294 out_unlock_not_found:
1300 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1302 * Within unwritten extents, the page cache determines which parts are holes
1303 * and which are data: uptodate buffer heads count as data; everything else
1306 * Returns the resulting offset on successs, and -ENOENT otherwise.
1309 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1312 pgoff_t index = offset >> PAGE_SHIFT;
1313 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1314 loff_t lastoff = offset;
1315 struct pagevec pvec;
1320 pagevec_init(&pvec);
1323 unsigned nr_pages, i;
1325 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1330 for (i = 0; i < nr_pages; i++) {
1331 struct page *page = pvec.pages[i];
1333 if (page_seek_hole_data(inode, page, &lastoff, whence))
1335 lastoff = page_offset(page) + PAGE_SIZE;
1337 pagevec_release(&pvec);
1338 } while (index < end);
1340 /* When no page at lastoff and we are not done, we found a hole. */
1341 if (whence != SEEK_HOLE)
1345 if (lastoff < offset + length)
1350 pagevec_release(&pvec);
1356 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1357 void *data, struct iomap *iomap)
1359 switch (iomap->type) {
1360 case IOMAP_UNWRITTEN:
1361 offset = page_cache_seek_hole_data(inode, offset, length,
1367 *(loff_t *)data = offset;
1375 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1377 loff_t size = i_size_read(inode);
1378 loff_t length = size - offset;
1381 /* Nothing to be found before or beyond the end of the file. */
1382 if (offset < 0 || offset >= size)
1385 while (length > 0) {
1386 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1387 &offset, iomap_seek_hole_actor);
1399 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1402 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1403 void *data, struct iomap *iomap)
1405 switch (iomap->type) {
1408 case IOMAP_UNWRITTEN:
1409 offset = page_cache_seek_hole_data(inode, offset, length,
1415 *(loff_t *)data = offset;
1421 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1423 loff_t size = i_size_read(inode);
1424 loff_t length = size - offset;
1427 /* Nothing to be found before or beyond the end of the file. */
1428 if (offset < 0 || offset >= size)
1431 while (length > 0) {
1432 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1433 &offset, iomap_seek_data_actor);
1447 EXPORT_SYMBOL_GPL(iomap_seek_data);
1450 * Private flags for iomap_dio, must not overlap with the public ones in
1453 #define IOMAP_DIO_WRITE_FUA (1 << 28)
1454 #define IOMAP_DIO_NEED_SYNC (1 << 29)
1455 #define IOMAP_DIO_WRITE (1 << 30)
1456 #define IOMAP_DIO_DIRTY (1 << 31)
1460 iomap_dio_end_io_t *end_io;
1466 bool wait_for_completion;
1469 /* used during submission and for synchronous completion: */
1471 struct iov_iter *iter;
1472 struct task_struct *waiter;
1473 struct request_queue *last_queue;
1477 /* used for aio completion: */
1479 struct work_struct work;
1484 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1486 struct kiocb *iocb = dio->iocb;
1487 struct inode *inode = file_inode(iocb->ki_filp);
1488 loff_t offset = iocb->ki_pos;
1492 ret = dio->end_io(iocb,
1493 dio->error ? dio->error : dio->size,
1501 /* check for short read */
1502 if (offset + ret > dio->i_size &&
1503 !(dio->flags & IOMAP_DIO_WRITE))
1504 ret = dio->i_size - offset;
1505 iocb->ki_pos += ret;
1509 * Try again to invalidate clean pages which might have been cached by
1510 * non-direct readahead, or faulted in by get_user_pages() if the source
1511 * of the write was an mmap'ed region of the file we're writing. Either
1512 * one is a pretty crazy thing to do, so we don't support it 100%. If
1513 * this invalidation fails, tough, the write still worked...
1515 * And this page cache invalidation has to be after dio->end_io(), as
1516 * some filesystems convert unwritten extents to real allocations in
1517 * end_io() when necessary, otherwise a racing buffer read would cache
1518 * zeros from unwritten extents.
1521 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1523 err = invalidate_inode_pages2_range(inode->i_mapping,
1524 offset >> PAGE_SHIFT,
1525 (offset + dio->size - 1) >> PAGE_SHIFT);
1527 dio_warn_stale_pagecache(iocb->ki_filp);
1531 * If this is a DSYNC write, make sure we push it to stable storage now
1532 * that we've written data.
1534 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1535 ret = generic_write_sync(iocb, ret);
1537 inode_dio_end(file_inode(iocb->ki_filp));
1543 static void iomap_dio_complete_work(struct work_struct *work)
1545 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1546 struct kiocb *iocb = dio->iocb;
1548 iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1552 * Set an error in the dio if none is set yet. We have to use cmpxchg
1553 * as the submission context and the completion context(s) can race to
1556 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1558 cmpxchg(&dio->error, 0, ret);
1561 static void iomap_dio_bio_end_io(struct bio *bio)
1563 struct iomap_dio *dio = bio->bi_private;
1564 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1567 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1569 if (atomic_dec_and_test(&dio->ref)) {
1570 if (dio->wait_for_completion) {
1571 struct task_struct *waiter = dio->submit.waiter;
1572 WRITE_ONCE(dio->submit.waiter, NULL);
1573 wake_up_process(waiter);
1574 } else if (dio->flags & IOMAP_DIO_WRITE) {
1575 struct inode *inode = file_inode(dio->iocb->ki_filp);
1577 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1578 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1580 iomap_dio_complete_work(&dio->aio.work);
1585 bio_check_pages_dirty(bio);
1587 struct bio_vec *bvec;
1590 bio_for_each_segment_all(bvec, bio, i)
1591 put_page(bvec->bv_page);
1597 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1600 struct page *page = ZERO_PAGE(0);
1603 bio = bio_alloc(GFP_KERNEL, 1);
1604 bio_set_dev(bio, iomap->bdev);
1605 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1606 bio->bi_private = dio;
1607 bio->bi_end_io = iomap_dio_bio_end_io;
1610 __bio_add_page(bio, page, len, 0);
1611 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
1613 atomic_inc(&dio->ref);
1614 return submit_bio(bio);
1618 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1619 struct iomap_dio *dio, struct iomap *iomap)
1621 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1622 unsigned int fs_block_size = i_blocksize(inode), pad;
1623 unsigned int align = iov_iter_alignment(dio->submit.iter);
1624 struct iov_iter iter;
1626 bool need_zeroout = false;
1627 bool use_fua = false;
1628 int nr_pages, ret = 0;
1631 if ((pos | length | align) & ((1 << blkbits) - 1))
1634 if (iomap->type == IOMAP_UNWRITTEN) {
1635 dio->flags |= IOMAP_DIO_UNWRITTEN;
1636 need_zeroout = true;
1639 if (iomap->flags & IOMAP_F_SHARED)
1640 dio->flags |= IOMAP_DIO_COW;
1642 if (iomap->flags & IOMAP_F_NEW) {
1643 need_zeroout = true;
1644 } else if (iomap->type == IOMAP_MAPPED) {
1646 * Use a FUA write if we need datasync semantics, this is a pure
1647 * data IO that doesn't require any metadata updates (including
1648 * after IO completion such as unwritten extent conversion) and
1649 * the underlying device supports FUA. This allows us to avoid
1650 * cache flushes on IO completion.
1652 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1653 (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1654 blk_queue_fua(bdev_get_queue(iomap->bdev)))
1659 * Operate on a partial iter trimmed to the extent we were called for.
1660 * We'll update the iter in the dio once we're done with this extent.
1662 iter = *dio->submit.iter;
1663 iov_iter_truncate(&iter, length);
1665 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1670 /* zero out from the start of the block to the write offset */
1671 pad = pos & (fs_block_size - 1);
1673 iomap_dio_zero(dio, iomap, pos - pad, pad);
1679 iov_iter_revert(dio->submit.iter, copied);
1683 bio = bio_alloc(GFP_KERNEL, nr_pages);
1684 bio_set_dev(bio, iomap->bdev);
1685 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1686 bio->bi_write_hint = dio->iocb->ki_hint;
1687 bio->bi_ioprio = dio->iocb->ki_ioprio;
1688 bio->bi_private = dio;
1689 bio->bi_end_io = iomap_dio_bio_end_io;
1691 ret = bio_iov_iter_get_pages(bio, &iter);
1692 if (unlikely(ret)) {
1694 * We have to stop part way through an IO. We must fall
1695 * through to the sub-block tail zeroing here, otherwise
1696 * this short IO may expose stale data in the tail of
1697 * the block we haven't written data to.
1703 n = bio->bi_iter.bi_size;
1704 if (dio->flags & IOMAP_DIO_WRITE) {
1705 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1707 bio->bi_opf |= REQ_FUA;
1709 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1710 task_io_account_write(n);
1712 bio->bi_opf = REQ_OP_READ;
1713 if (dio->flags & IOMAP_DIO_DIRTY)
1714 bio_set_pages_dirty(bio);
1717 iov_iter_advance(dio->submit.iter, n);
1723 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1725 atomic_inc(&dio->ref);
1727 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1728 dio->submit.cookie = submit_bio(bio);
1732 * We need to zeroout the tail of a sub-block write if the extent type
1733 * requires zeroing or the write extends beyond EOF. If we don't zero
1734 * the block tail in the latter case, we can expose stale data via mmap
1735 * reads of the EOF block.
1739 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
1740 /* zero out from the end of the write to the end of the block */
1741 pad = pos & (fs_block_size - 1);
1743 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1745 return copied ? copied : ret;
1749 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1751 length = iov_iter_zero(length, dio->submit.iter);
1752 dio->size += length;
1757 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1758 struct iomap_dio *dio, struct iomap *iomap)
1760 struct iov_iter *iter = dio->submit.iter;
1763 BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1765 if (dio->flags & IOMAP_DIO_WRITE) {
1766 loff_t size = inode->i_size;
1769 memset(iomap->inline_data + size, 0, pos - size);
1770 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1772 if (pos + copied > size)
1773 i_size_write(inode, pos + copied);
1774 mark_inode_dirty(inode);
1777 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1779 dio->size += copied;
1784 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1785 void *data, struct iomap *iomap)
1787 struct iomap_dio *dio = data;
1789 switch (iomap->type) {
1791 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1793 return iomap_dio_hole_actor(length, dio);
1794 case IOMAP_UNWRITTEN:
1795 if (!(dio->flags & IOMAP_DIO_WRITE))
1796 return iomap_dio_hole_actor(length, dio);
1797 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1799 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1801 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1809 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1810 * is being issued as AIO or not. This allows us to optimise pure data writes
1811 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1812 * REQ_FLUSH post write. This is slightly tricky because a single request here
1813 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1814 * may be pure data writes. In that case, we still need to do a full data sync
1818 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1819 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1821 struct address_space *mapping = iocb->ki_filp->f_mapping;
1822 struct inode *inode = file_inode(iocb->ki_filp);
1823 size_t count = iov_iter_count(iter);
1824 loff_t pos = iocb->ki_pos, start = pos;
1825 loff_t end = iocb->ki_pos + count - 1, ret = 0;
1826 unsigned int flags = IOMAP_DIRECT;
1827 bool wait_for_completion = is_sync_kiocb(iocb);
1828 struct blk_plug plug;
1829 struct iomap_dio *dio;
1831 lockdep_assert_held(&inode->i_rwsem);
1836 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1841 atomic_set(&dio->ref, 1);
1843 dio->i_size = i_size_read(inode);
1844 dio->end_io = end_io;
1848 dio->submit.iter = iter;
1849 dio->submit.waiter = current;
1850 dio->submit.cookie = BLK_QC_T_NONE;
1851 dio->submit.last_queue = NULL;
1853 if (iov_iter_rw(iter) == READ) {
1854 if (pos >= dio->i_size)
1857 if (iter->type == ITER_IOVEC)
1858 dio->flags |= IOMAP_DIO_DIRTY;
1860 flags |= IOMAP_WRITE;
1861 dio->flags |= IOMAP_DIO_WRITE;
1863 /* for data sync or sync, we need sync completion processing */
1864 if (iocb->ki_flags & IOCB_DSYNC)
1865 dio->flags |= IOMAP_DIO_NEED_SYNC;
1868 * For datasync only writes, we optimistically try using FUA for
1869 * this IO. Any non-FUA write that occurs will clear this flag,
1870 * hence we know before completion whether a cache flush is
1873 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1874 dio->flags |= IOMAP_DIO_WRITE_FUA;
1877 if (iocb->ki_flags & IOCB_NOWAIT) {
1878 if (filemap_range_has_page(mapping, start, end)) {
1882 flags |= IOMAP_NOWAIT;
1885 ret = filemap_write_and_wait_range(mapping, start, end);
1890 * Try to invalidate cache pages for the range we're direct
1891 * writing. If this invalidation fails, tough, the write will
1892 * still work, but racing two incompatible write paths is a
1893 * pretty crazy thing to do, so we don't support it 100%.
1895 ret = invalidate_inode_pages2_range(mapping,
1896 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1898 dio_warn_stale_pagecache(iocb->ki_filp);
1901 if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1902 !inode->i_sb->s_dio_done_wq) {
1903 ret = sb_init_dio_done_wq(inode->i_sb);
1908 inode_dio_begin(inode);
1910 blk_start_plug(&plug);
1912 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1915 /* magic error code to fall back to buffered I/O */
1916 if (ret == -ENOTBLK) {
1917 wait_for_completion = true;
1924 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
1926 * We only report that we've read data up to i_size.
1927 * Revert iter to a state corresponding to that as
1928 * some callers (such as splice code) rely on it.
1930 iov_iter_revert(iter, pos - dio->i_size);
1933 } while ((count = iov_iter_count(iter)) > 0);
1934 blk_finish_plug(&plug);
1937 iomap_dio_set_error(dio, ret);
1940 * If all the writes we issued were FUA, we don't need to flush the
1941 * cache on IO completion. Clear the sync flag for this case.
1943 if (dio->flags & IOMAP_DIO_WRITE_FUA)
1944 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1947 * We are about to drop our additional submission reference, which
1948 * might be the last reference to the dio. There are three three
1949 * different ways we can progress here:
1951 * (a) If this is the last reference we will always complete and free
1952 * the dio ourselves.
1953 * (b) If this is not the last reference, and we serve an asynchronous
1954 * iocb, we must never touch the dio after the decrement, the
1955 * I/O completion handler will complete and free it.
1956 * (c) If this is not the last reference, but we serve a synchronous
1957 * iocb, the I/O completion handler will wake us up on the drop
1958 * of the final reference, and we will complete and free it here
1959 * after we got woken by the I/O completion handler.
1961 dio->wait_for_completion = wait_for_completion;
1962 if (!atomic_dec_and_test(&dio->ref)) {
1963 if (!wait_for_completion)
1964 return -EIOCBQUEUED;
1967 set_current_state(TASK_UNINTERRUPTIBLE);
1968 if (!READ_ONCE(dio->submit.waiter))
1971 if (!(iocb->ki_flags & IOCB_HIPRI) ||
1972 !dio->submit.last_queue ||
1973 !blk_poll(dio->submit.last_queue,
1974 dio->submit.cookie))
1977 __set_current_state(TASK_RUNNING);
1980 return iomap_dio_complete(dio);
1986 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1988 /* Swapfile activation */
1991 struct iomap_swapfile_info {
1992 struct iomap iomap; /* accumulated iomap */
1993 struct swap_info_struct *sis;
1994 uint64_t lowest_ppage; /* lowest physical addr seen (pages) */
1995 uint64_t highest_ppage; /* highest physical addr seen (pages) */
1996 unsigned long nr_pages; /* number of pages collected */
1997 int nr_extents; /* extent count */
2001 * Collect physical extents for this swap file. Physical extents reported to
2002 * the swap code must be trimmed to align to a page boundary. The logical
2003 * offset within the file is irrelevant since the swapfile code maps logical
2004 * page numbers of the swap device to the physical page-aligned extents.
2006 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
2008 struct iomap *iomap = &isi->iomap;
2009 unsigned long nr_pages;
2010 uint64_t first_ppage;
2011 uint64_t first_ppage_reported;
2012 uint64_t next_ppage;
2016 * Round the start up and the end down so that the physical
2017 * extent aligns to a page boundary.
2019 first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
2020 next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
2023 /* Skip too-short physical extents. */
2024 if (first_ppage >= next_ppage)
2026 nr_pages = next_ppage - first_ppage;
2029 * Calculate how much swap space we're adding; the first page contains
2030 * the swap header and doesn't count. The mm still wants that first
2031 * page fed to add_swap_extent, however.
2033 first_ppage_reported = first_ppage;
2034 if (iomap->offset == 0)
2035 first_ppage_reported++;
2036 if (isi->lowest_ppage > first_ppage_reported)
2037 isi->lowest_ppage = first_ppage_reported;
2038 if (isi->highest_ppage < (next_ppage - 1))
2039 isi->highest_ppage = next_ppage - 1;
2041 /* Add extent, set up for the next call. */
2042 error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
2045 isi->nr_extents += error;
2046 isi->nr_pages += nr_pages;
2051 * Accumulate iomaps for this swap file. We have to accumulate iomaps because
2052 * swap only cares about contiguous page-aligned physical extents and makes no
2053 * distinction between written and unwritten extents.
2055 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2056 loff_t count, void *data, struct iomap *iomap)
2058 struct iomap_swapfile_info *isi = data;
2061 switch (iomap->type) {
2063 case IOMAP_UNWRITTEN:
2064 /* Only real or unwritten extents. */
2067 /* No inline data. */
2068 pr_err("swapon: file is inline\n");
2071 pr_err("swapon: file has unallocated extents\n");
2075 /* No uncommitted metadata or shared blocks. */
2076 if (iomap->flags & IOMAP_F_DIRTY) {
2077 pr_err("swapon: file is not committed\n");
2080 if (iomap->flags & IOMAP_F_SHARED) {
2081 pr_err("swapon: file has shared extents\n");
2085 /* Only one bdev per swap file. */
2086 if (iomap->bdev != isi->sis->bdev) {
2087 pr_err("swapon: file is on multiple devices\n");
2091 if (isi->iomap.length == 0) {
2092 /* No accumulated extent, so just store it. */
2093 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2094 } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2095 /* Append this to the accumulated extent. */
2096 isi->iomap.length += iomap->length;
2098 /* Otherwise, add the retained iomap and store this one. */
2099 error = iomap_swapfile_add_extent(isi);
2102 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2108 * Iterate a swap file's iomaps to construct physical extents that can be
2109 * passed to the swapfile subsystem.
2111 int iomap_swapfile_activate(struct swap_info_struct *sis,
2112 struct file *swap_file, sector_t *pagespan,
2113 const struct iomap_ops *ops)
2115 struct iomap_swapfile_info isi = {
2117 .lowest_ppage = (sector_t)-1ULL,
2119 struct address_space *mapping = swap_file->f_mapping;
2120 struct inode *inode = mapping->host;
2122 loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2126 * Persist all file mapping metadata so that we won't have any
2127 * IOMAP_F_DIRTY iomaps.
2129 ret = vfs_fsync(swap_file, 1);
2134 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2135 ops, &isi, iomap_swapfile_activate_actor);
2143 if (isi.iomap.length) {
2144 ret = iomap_swapfile_add_extent(&isi);
2149 *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2150 sis->max = isi.nr_pages;
2151 sis->pages = isi.nr_pages - 1;
2152 sis->highest_bit = isi.nr_pages - 1;
2153 return isi.nr_extents;
2155 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2156 #endif /* CONFIG_SWAP */
2159 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2160 void *data, struct iomap *iomap)
2162 sector_t *bno = data, addr;
2164 if (iomap->type == IOMAP_MAPPED) {
2165 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2167 WARN(1, "would truncate bmap result\n");
2174 /* legacy ->bmap interface. 0 is the error return (!) */
2176 iomap_bmap(struct address_space *mapping, sector_t bno,
2177 const struct iomap_ops *ops)
2179 struct inode *inode = mapping->host;
2180 loff_t pos = bno << inode->i_blkbits;
2181 unsigned blocksize = i_blocksize(inode);
2183 if (filemap_write_and_wait(mapping))
2187 iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2190 EXPORT_SYMBOL_GPL(iomap_bmap);