2 * Copyright (C) 2010 Red Hat, Inc.
3 * Copyright (c) 2016 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>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/pagevec.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/backing-dev.h>
27 #include <linux/buffer_head.h>
28 #include <linux/task_io_accounting_ops.h>
29 #include <linux/dax.h>
30 #include <linux/sched/signal.h>
31 #include <linux/swap.h>
36 * Execute a iomap write on a segment of the mapping that spans a
37 * contiguous range of pages that have identical block mapping state.
39 * This avoids the need to map pages individually, do individual allocations
40 * for each page and most importantly avoid the need for filesystem specific
41 * locking per page. Instead, all the operations are amortised over the entire
42 * range of pages. It is assumed that the filesystems will lock whatever
43 * resources they require in the iomap_begin call, and release them in the
47 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
48 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
50 struct iomap iomap = { 0 };
51 loff_t written = 0, ret;
54 * Need to map a range from start position for length bytes. This can
55 * span multiple pages - it is only guaranteed to return a range of a
56 * single type of pages (e.g. all into a hole, all mapped or all
57 * unwritten). Failure at this point has nothing to undo.
59 * If allocation is required for this range, reserve the space now so
60 * that the allocation is guaranteed to succeed later on. Once we copy
61 * the data into the page cache pages, then we cannot fail otherwise we
62 * expose transient stale data. If the reserve fails, we can safely
63 * back out at this point as there is nothing to undo.
65 ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
68 if (WARN_ON(iomap.offset > pos))
70 if (WARN_ON(iomap.length == 0))
74 * Cut down the length to the one actually provided by the filesystem,
75 * as it might not be able to give us the whole size that we requested.
77 if (iomap.offset + iomap.length < pos + length)
78 length = iomap.offset + iomap.length - pos;
81 * Now that we have guaranteed that the space allocation will succeed.
82 * we can do the copy-in page by page without having to worry about
83 * failures exposing transient data.
85 written = actor(inode, pos, length, data, &iomap);
88 * Now the data has been copied, commit the range we've copied. This
89 * should not fail unless the filesystem has had a fatal error.
92 ret = ops->iomap_end(inode, pos, length,
93 written > 0 ? written : 0,
97 return written ? written : ret;
101 iomap_sector(struct iomap *iomap, loff_t pos)
103 return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
107 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
109 loff_t i_size = i_size_read(inode);
112 * Only truncate newly allocated pages beyoned EOF, even if the
113 * write started inside the existing inode size.
115 if (pos + len > i_size)
116 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
120 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
121 struct page **pagep, struct iomap *iomap)
123 pgoff_t index = pos >> PAGE_SHIFT;
127 BUG_ON(pos + len > iomap->offset + iomap->length);
129 if (fatal_signal_pending(current))
132 page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
136 status = __block_write_begin_int(page, pos, len, NULL, iomap);
137 if (unlikely(status)) {
142 iomap_write_failed(inode, pos, len);
150 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
151 unsigned copied, struct page *page)
155 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
158 iomap_write_failed(inode, pos, len);
163 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
166 struct iov_iter *i = data;
169 unsigned int flags = AOP_FLAG_NOFS;
173 unsigned long offset; /* Offset into pagecache page */
174 unsigned long bytes; /* Bytes to write to page */
175 size_t copied; /* Bytes copied from user */
177 offset = (pos & (PAGE_SIZE - 1));
178 bytes = min_t(unsigned long, PAGE_SIZE - offset,
185 * Bring in the user page that we will copy from _first_.
186 * Otherwise there's a nasty deadlock on copying from the
187 * same page as we're writing to, without it being marked
190 * Not only is this an optimisation, but it is also required
191 * to check that the address is actually valid, when atomic
192 * usercopies are used, below.
194 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
199 status = iomap_write_begin(inode, pos, bytes, flags, &page,
201 if (unlikely(status))
204 if (mapping_writably_mapped(inode->i_mapping))
205 flush_dcache_page(page);
207 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
209 flush_dcache_page(page);
211 status = iomap_write_end(inode, pos, bytes, copied, page);
212 if (unlikely(status < 0))
218 iov_iter_advance(i, copied);
219 if (unlikely(copied == 0)) {
221 * If we were unable to copy any data at all, we must
222 * fall back to a single segment length write.
224 * If we didn't fallback here, we could livelock
225 * because not all segments in the iov can be copied at
226 * once without a pagefault.
228 bytes = min_t(unsigned long, PAGE_SIZE - offset,
229 iov_iter_single_seg_count(i));
236 balance_dirty_pages_ratelimited(inode->i_mapping);
237 } while (iov_iter_count(i) && length);
239 return written ? written : status;
243 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
244 const struct iomap_ops *ops)
246 struct inode *inode = iocb->ki_filp->f_mapping->host;
247 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
249 while (iov_iter_count(iter)) {
250 ret = iomap_apply(inode, pos, iov_iter_count(iter),
251 IOMAP_WRITE, ops, iter, iomap_write_actor);
258 return written ? written : ret;
260 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
263 __iomap_read_page(struct inode *inode, loff_t offset)
265 struct address_space *mapping = inode->i_mapping;
268 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
271 if (!PageUptodate(page)) {
273 return ERR_PTR(-EIO);
279 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
286 struct page *page, *rpage;
287 unsigned long offset; /* Offset into pagecache page */
288 unsigned long bytes; /* Bytes to write to page */
290 offset = (pos & (PAGE_SIZE - 1));
291 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
293 rpage = __iomap_read_page(inode, pos);
295 return PTR_ERR(rpage);
297 status = iomap_write_begin(inode, pos, bytes,
298 AOP_FLAG_NOFS, &page, iomap);
300 if (unlikely(status))
303 WARN_ON_ONCE(!PageUptodate(page));
305 status = iomap_write_end(inode, pos, bytes, bytes, page);
306 if (unlikely(status <= 0)) {
307 if (WARN_ON_ONCE(status == 0))
318 balance_dirty_pages_ratelimited(inode->i_mapping);
325 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
326 const struct iomap_ops *ops)
331 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
341 EXPORT_SYMBOL_GPL(iomap_file_dirty);
343 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
344 unsigned bytes, struct iomap *iomap)
349 status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
354 zero_user(page, offset, bytes);
355 mark_page_accessed(page);
357 return iomap_write_end(inode, pos, bytes, bytes, page);
360 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
363 return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
364 iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
368 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
369 void *data, struct iomap *iomap)
371 bool *did_zero = data;
375 /* already zeroed? we're done. */
376 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
380 unsigned offset, bytes;
382 offset = pos & (PAGE_SIZE - 1); /* Within page */
383 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
386 status = iomap_dax_zero(pos, offset, bytes, iomap);
388 status = iomap_zero(inode, pos, offset, bytes, iomap);
403 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
404 const struct iomap_ops *ops)
409 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
410 ops, did_zero, iomap_zero_range_actor);
420 EXPORT_SYMBOL_GPL(iomap_zero_range);
423 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
424 const struct iomap_ops *ops)
426 unsigned int blocksize = i_blocksize(inode);
427 unsigned int off = pos & (blocksize - 1);
429 /* Block boundary? Nothing to do */
432 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
434 EXPORT_SYMBOL_GPL(iomap_truncate_page);
437 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
438 void *data, struct iomap *iomap)
440 struct page *page = data;
443 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
447 block_commit_write(page, 0, length);
451 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
453 struct page *page = vmf->page;
454 struct inode *inode = file_inode(vmf->vma->vm_file);
455 unsigned long length;
460 size = i_size_read(inode);
461 if ((page->mapping != inode->i_mapping) ||
462 (page_offset(page) > size)) {
463 /* We overload EFAULT to mean page got truncated */
468 /* page is wholly or partially inside EOF */
469 if (((page->index + 1) << PAGE_SHIFT) > size)
470 length = size & ~PAGE_MASK;
474 offset = page_offset(page);
476 ret = iomap_apply(inode, offset, length,
477 IOMAP_WRITE | IOMAP_FAULT, ops, page,
478 iomap_page_mkwrite_actor);
479 if (unlikely(ret <= 0))
485 set_page_dirty(page);
486 wait_for_stable_page(page);
487 return VM_FAULT_LOCKED;
490 return block_page_mkwrite_return(ret);
492 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
495 struct fiemap_extent_info *fi;
499 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
500 struct iomap *iomap, u32 flags)
502 switch (iomap->type) {
507 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
511 case IOMAP_UNWRITTEN:
512 flags |= FIEMAP_EXTENT_UNWRITTEN;
515 flags |= FIEMAP_EXTENT_DATA_INLINE;
519 if (iomap->flags & IOMAP_F_MERGED)
520 flags |= FIEMAP_EXTENT_MERGED;
521 if (iomap->flags & IOMAP_F_SHARED)
522 flags |= FIEMAP_EXTENT_SHARED;
524 return fiemap_fill_next_extent(fi, iomap->offset,
525 iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
526 iomap->length, flags);
530 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
533 struct fiemap_ctx *ctx = data;
536 if (iomap->type == IOMAP_HOLE)
539 ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
542 case 0: /* success */
544 case 1: /* extent array full */
551 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
552 loff_t start, loff_t len, const struct iomap_ops *ops)
554 struct fiemap_ctx ctx;
557 memset(&ctx, 0, sizeof(ctx));
559 ctx.prev.type = IOMAP_HOLE;
561 ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
565 if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
566 ret = filemap_write_and_wait(inode->i_mapping);
572 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
574 /* inode with no (attribute) mapping will give ENOENT */
586 if (ctx.prev.type != IOMAP_HOLE) {
587 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
594 EXPORT_SYMBOL_GPL(iomap_fiemap);
597 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
598 * Returns true if found and updates @lastoff to the offset in file.
601 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
604 const struct address_space_operations *ops = inode->i_mapping->a_ops;
605 unsigned int bsize = i_blocksize(inode), off;
606 bool seek_data = whence == SEEK_DATA;
607 loff_t poff = page_offset(page);
609 if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
612 if (*lastoff < poff) {
614 * Last offset smaller than the start of the page means we found
617 if (whence == SEEK_HOLE)
623 * Just check the page unless we can and should check block ranges:
625 if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
626 return PageUptodate(page) == seek_data;
629 if (unlikely(page->mapping != inode->i_mapping))
630 goto out_unlock_not_found;
632 for (off = 0; off < PAGE_SIZE; off += bsize) {
633 if ((*lastoff & ~PAGE_MASK) >= off + bsize)
635 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
639 *lastoff = poff + off + bsize;
642 out_unlock_not_found:
648 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
650 * Within unwritten extents, the page cache determines which parts are holes
651 * and which are data: uptodate buffer heads count as data; everything else
654 * Returns the resulting offset on successs, and -ENOENT otherwise.
657 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
660 pgoff_t index = offset >> PAGE_SHIFT;
661 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
662 loff_t lastoff = offset;
671 unsigned nr_pages, i;
673 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
678 for (i = 0; i < nr_pages; i++) {
679 struct page *page = pvec.pages[i];
681 if (page_seek_hole_data(inode, page, &lastoff, whence))
683 lastoff = page_offset(page) + PAGE_SIZE;
685 pagevec_release(&pvec);
686 } while (index < end);
688 /* When no page at lastoff and we are not done, we found a hole. */
689 if (whence != SEEK_HOLE)
693 if (lastoff < offset + length)
698 pagevec_release(&pvec);
704 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
705 void *data, struct iomap *iomap)
707 switch (iomap->type) {
708 case IOMAP_UNWRITTEN:
709 offset = page_cache_seek_hole_data(inode, offset, length,
715 *(loff_t *)data = offset;
723 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
725 loff_t size = i_size_read(inode);
726 loff_t length = size - offset;
729 /* Nothing to be found before or beyond the end of the file. */
730 if (offset < 0 || offset >= size)
734 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
735 &offset, iomap_seek_hole_actor);
747 EXPORT_SYMBOL_GPL(iomap_seek_hole);
750 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
751 void *data, struct iomap *iomap)
753 switch (iomap->type) {
756 case IOMAP_UNWRITTEN:
757 offset = page_cache_seek_hole_data(inode, offset, length,
763 *(loff_t *)data = offset;
769 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
771 loff_t size = i_size_read(inode);
772 loff_t length = size - offset;
775 /* Nothing to be found before or beyond the end of the file. */
776 if (offset < 0 || offset >= size)
780 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
781 &offset, iomap_seek_data_actor);
795 EXPORT_SYMBOL_GPL(iomap_seek_data);
798 * Private flags for iomap_dio, must not overlap with the public ones in
801 #define IOMAP_DIO_WRITE_FUA (1 << 28)
802 #define IOMAP_DIO_NEED_SYNC (1 << 29)
803 #define IOMAP_DIO_WRITE (1 << 30)
804 #define IOMAP_DIO_DIRTY (1 << 31)
808 iomap_dio_end_io_t *end_io;
816 /* used during submission and for synchronous completion: */
818 struct iov_iter *iter;
819 struct task_struct *waiter;
820 struct request_queue *last_queue;
824 /* used for aio completion: */
826 struct work_struct work;
831 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
833 struct kiocb *iocb = dio->iocb;
834 struct inode *inode = file_inode(iocb->ki_filp);
835 loff_t offset = iocb->ki_pos;
839 ret = dio->end_io(iocb,
840 dio->error ? dio->error : dio->size,
848 /* check for short read */
849 if (offset + ret > dio->i_size &&
850 !(dio->flags & IOMAP_DIO_WRITE))
851 ret = dio->i_size - offset;
856 * Try again to invalidate clean pages which might have been cached by
857 * non-direct readahead, or faulted in by get_user_pages() if the source
858 * of the write was an mmap'ed region of the file we're writing. Either
859 * one is a pretty crazy thing to do, so we don't support it 100%. If
860 * this invalidation fails, tough, the write still worked...
862 * And this page cache invalidation has to be after dio->end_io(), as
863 * some filesystems convert unwritten extents to real allocations in
864 * end_io() when necessary, otherwise a racing buffer read would cache
865 * zeros from unwritten extents.
868 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
870 err = invalidate_inode_pages2_range(inode->i_mapping,
871 offset >> PAGE_SHIFT,
872 (offset + dio->size - 1) >> PAGE_SHIFT);
874 dio_warn_stale_pagecache(iocb->ki_filp);
878 * If this is a DSYNC write, make sure we push it to stable storage now
879 * that we've written data.
881 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
882 ret = generic_write_sync(iocb, ret);
884 inode_dio_end(file_inode(iocb->ki_filp));
890 static void iomap_dio_complete_work(struct work_struct *work)
892 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
893 struct kiocb *iocb = dio->iocb;
895 iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
899 * Set an error in the dio if none is set yet. We have to use cmpxchg
900 * as the submission context and the completion context(s) can race to
903 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
905 cmpxchg(&dio->error, 0, ret);
908 static void iomap_dio_bio_end_io(struct bio *bio)
910 struct iomap_dio *dio = bio->bi_private;
911 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
914 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
916 if (atomic_dec_and_test(&dio->ref)) {
917 if (is_sync_kiocb(dio->iocb)) {
918 struct task_struct *waiter = dio->submit.waiter;
920 WRITE_ONCE(dio->submit.waiter, NULL);
921 wake_up_process(waiter);
922 } else if (dio->flags & IOMAP_DIO_WRITE) {
923 struct inode *inode = file_inode(dio->iocb->ki_filp);
925 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
926 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
928 iomap_dio_complete_work(&dio->aio.work);
933 bio_check_pages_dirty(bio);
935 struct bio_vec *bvec;
938 bio_for_each_segment_all(bvec, bio, i)
939 put_page(bvec->bv_page);
945 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
948 struct page *page = ZERO_PAGE(0);
951 bio = bio_alloc(GFP_KERNEL, 1);
952 bio_set_dev(bio, iomap->bdev);
953 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
954 bio->bi_private = dio;
955 bio->bi_end_io = iomap_dio_bio_end_io;
958 __bio_add_page(bio, page, len, 0);
959 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
961 atomic_inc(&dio->ref);
962 return submit_bio(bio);
966 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
967 void *data, struct iomap *iomap)
969 struct iomap_dio *dio = data;
970 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
971 unsigned int fs_block_size = i_blocksize(inode), pad;
972 unsigned int align = iov_iter_alignment(dio->submit.iter);
973 struct iov_iter iter;
975 bool need_zeroout = false;
976 bool use_fua = false;
980 if ((pos | length | align) & ((1 << blkbits) - 1))
983 switch (iomap->type) {
985 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
988 case IOMAP_UNWRITTEN:
989 if (!(dio->flags & IOMAP_DIO_WRITE)) {
990 length = iov_iter_zero(length, dio->submit.iter);
994 dio->flags |= IOMAP_DIO_UNWRITTEN;
998 if (iomap->flags & IOMAP_F_SHARED)
999 dio->flags |= IOMAP_DIO_COW;
1000 if (iomap->flags & IOMAP_F_NEW) {
1001 need_zeroout = true;
1004 * Use a FUA write if we need datasync semantics, this
1005 * is a pure data IO that doesn't require any metadata
1006 * updates and the underlying device supports FUA. This
1007 * allows us to avoid cache flushes on IO completion.
1009 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1010 (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1011 blk_queue_fua(bdev_get_queue(iomap->bdev)))
1021 * Operate on a partial iter trimmed to the extent we were called for.
1022 * We'll update the iter in the dio once we're done with this extent.
1024 iter = *dio->submit.iter;
1025 iov_iter_truncate(&iter, length);
1027 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1032 /* zero out from the start of the block to the write offset */
1033 pad = pos & (fs_block_size - 1);
1035 iomap_dio_zero(dio, iomap, pos - pad, pad);
1041 iov_iter_revert(dio->submit.iter, copied);
1045 bio = bio_alloc(GFP_KERNEL, nr_pages);
1046 bio_set_dev(bio, iomap->bdev);
1047 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1048 bio->bi_write_hint = dio->iocb->ki_hint;
1049 bio->bi_private = dio;
1050 bio->bi_end_io = iomap_dio_bio_end_io;
1052 ret = bio_iov_iter_get_pages(bio, &iter);
1053 if (unlikely(ret)) {
1055 return copied ? copied : ret;
1058 n = bio->bi_iter.bi_size;
1059 if (dio->flags & IOMAP_DIO_WRITE) {
1060 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1062 bio->bi_opf |= REQ_FUA;
1064 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1065 task_io_account_write(n);
1067 bio->bi_opf = REQ_OP_READ;
1068 if (dio->flags & IOMAP_DIO_DIRTY)
1069 bio_set_pages_dirty(bio);
1072 iov_iter_advance(dio->submit.iter, n);
1078 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1080 atomic_inc(&dio->ref);
1082 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1083 dio->submit.cookie = submit_bio(bio);
1087 /* zero out from the end of the write to the end of the block */
1088 pad = pos & (fs_block_size - 1);
1090 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1096 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1097 * is being issued as AIO or not. This allows us to optimise pure data writes
1098 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1099 * REQ_FLUSH post write. This is slightly tricky because a single request here
1100 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1101 * may be pure data writes. In that case, we still need to do a full data sync
1105 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1106 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1108 struct address_space *mapping = iocb->ki_filp->f_mapping;
1109 struct inode *inode = file_inode(iocb->ki_filp);
1110 size_t count = iov_iter_count(iter);
1111 loff_t pos = iocb->ki_pos, start = pos;
1112 loff_t end = iocb->ki_pos + count - 1, ret = 0;
1113 unsigned int flags = IOMAP_DIRECT;
1114 struct blk_plug plug;
1115 struct iomap_dio *dio;
1117 lockdep_assert_held(&inode->i_rwsem);
1122 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1127 atomic_set(&dio->ref, 1);
1129 dio->i_size = i_size_read(inode);
1130 dio->end_io = end_io;
1134 dio->submit.iter = iter;
1135 if (is_sync_kiocb(iocb)) {
1136 dio->submit.waiter = current;
1137 dio->submit.cookie = BLK_QC_T_NONE;
1138 dio->submit.last_queue = NULL;
1141 if (iov_iter_rw(iter) == READ) {
1142 if (pos >= dio->i_size)
1145 if (iter->type == ITER_IOVEC)
1146 dio->flags |= IOMAP_DIO_DIRTY;
1148 flags |= IOMAP_WRITE;
1149 dio->flags |= IOMAP_DIO_WRITE;
1151 /* for data sync or sync, we need sync completion processing */
1152 if (iocb->ki_flags & IOCB_DSYNC)
1153 dio->flags |= IOMAP_DIO_NEED_SYNC;
1156 * For datasync only writes, we optimistically try using FUA for
1157 * this IO. Any non-FUA write that occurs will clear this flag,
1158 * hence we know before completion whether a cache flush is
1161 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1162 dio->flags |= IOMAP_DIO_WRITE_FUA;
1165 if (iocb->ki_flags & IOCB_NOWAIT) {
1166 if (filemap_range_has_page(mapping, start, end)) {
1170 flags |= IOMAP_NOWAIT;
1173 ret = filemap_write_and_wait_range(mapping, start, end);
1178 * Try to invalidate cache pages for the range we're direct
1179 * writing. If this invalidation fails, tough, the write will
1180 * still work, but racing two incompatible write paths is a
1181 * pretty crazy thing to do, so we don't support it 100%.
1183 ret = invalidate_inode_pages2_range(mapping,
1184 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1186 dio_warn_stale_pagecache(iocb->ki_filp);
1189 if (iov_iter_rw(iter) == WRITE && !is_sync_kiocb(iocb) &&
1190 !inode->i_sb->s_dio_done_wq) {
1191 ret = sb_init_dio_done_wq(inode->i_sb);
1196 inode_dio_begin(inode);
1198 blk_start_plug(&plug);
1200 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1203 /* magic error code to fall back to buffered I/O */
1204 if (ret == -ENOTBLK)
1210 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1212 } while ((count = iov_iter_count(iter)) > 0);
1213 blk_finish_plug(&plug);
1216 iomap_dio_set_error(dio, ret);
1219 * If all the writes we issued were FUA, we don't need to flush the
1220 * cache on IO completion. Clear the sync flag for this case.
1222 if (dio->flags & IOMAP_DIO_WRITE_FUA)
1223 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1225 if (!atomic_dec_and_test(&dio->ref)) {
1226 if (!is_sync_kiocb(iocb))
1227 return -EIOCBQUEUED;
1230 set_current_state(TASK_UNINTERRUPTIBLE);
1231 if (!READ_ONCE(dio->submit.waiter))
1234 if (!(iocb->ki_flags & IOCB_HIPRI) ||
1235 !dio->submit.last_queue ||
1236 !blk_poll(dio->submit.last_queue,
1237 dio->submit.cookie))
1240 __set_current_state(TASK_RUNNING);
1243 ret = iomap_dio_complete(dio);
1251 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1253 /* Swapfile activation */
1256 struct iomap_swapfile_info {
1257 struct iomap iomap; /* accumulated iomap */
1258 struct swap_info_struct *sis;
1259 uint64_t lowest_ppage; /* lowest physical addr seen (pages) */
1260 uint64_t highest_ppage; /* highest physical addr seen (pages) */
1261 unsigned long nr_pages; /* number of pages collected */
1262 int nr_extents; /* extent count */
1266 * Collect physical extents for this swap file. Physical extents reported to
1267 * the swap code must be trimmed to align to a page boundary. The logical
1268 * offset within the file is irrelevant since the swapfile code maps logical
1269 * page numbers of the swap device to the physical page-aligned extents.
1271 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
1273 struct iomap *iomap = &isi->iomap;
1274 unsigned long nr_pages;
1275 uint64_t first_ppage;
1276 uint64_t first_ppage_reported;
1277 uint64_t next_ppage;
1281 * Round the start up and the end down so that the physical
1282 * extent aligns to a page boundary.
1284 first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
1285 next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
1288 /* Skip too-short physical extents. */
1289 if (first_ppage >= next_ppage)
1291 nr_pages = next_ppage - first_ppage;
1294 * Calculate how much swap space we're adding; the first page contains
1295 * the swap header and doesn't count. The mm still wants that first
1296 * page fed to add_swap_extent, however.
1298 first_ppage_reported = first_ppage;
1299 if (iomap->offset == 0)
1300 first_ppage_reported++;
1301 if (isi->lowest_ppage > first_ppage_reported)
1302 isi->lowest_ppage = first_ppage_reported;
1303 if (isi->highest_ppage < (next_ppage - 1))
1304 isi->highest_ppage = next_ppage - 1;
1306 /* Add extent, set up for the next call. */
1307 error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
1310 isi->nr_extents += error;
1311 isi->nr_pages += nr_pages;
1316 * Accumulate iomaps for this swap file. We have to accumulate iomaps because
1317 * swap only cares about contiguous page-aligned physical extents and makes no
1318 * distinction between written and unwritten extents.
1320 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
1321 loff_t count, void *data, struct iomap *iomap)
1323 struct iomap_swapfile_info *isi = data;
1326 switch (iomap->type) {
1328 case IOMAP_UNWRITTEN:
1329 /* Only real or unwritten extents. */
1332 /* No inline data. */
1333 pr_err("swapon: file is inline\n");
1336 pr_err("swapon: file has unallocated extents\n");
1340 /* No uncommitted metadata or shared blocks. */
1341 if (iomap->flags & IOMAP_F_DIRTY) {
1342 pr_err("swapon: file is not committed\n");
1345 if (iomap->flags & IOMAP_F_SHARED) {
1346 pr_err("swapon: file has shared extents\n");
1350 /* Only one bdev per swap file. */
1351 if (iomap->bdev != isi->sis->bdev) {
1352 pr_err("swapon: file is on multiple devices\n");
1356 if (isi->iomap.length == 0) {
1357 /* No accumulated extent, so just store it. */
1358 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
1359 } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
1360 /* Append this to the accumulated extent. */
1361 isi->iomap.length += iomap->length;
1363 /* Otherwise, add the retained iomap and store this one. */
1364 error = iomap_swapfile_add_extent(isi);
1367 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
1373 * Iterate a swap file's iomaps to construct physical extents that can be
1374 * passed to the swapfile subsystem.
1376 int iomap_swapfile_activate(struct swap_info_struct *sis,
1377 struct file *swap_file, sector_t *pagespan,
1378 const struct iomap_ops *ops)
1380 struct iomap_swapfile_info isi = {
1382 .lowest_ppage = (sector_t)-1ULL,
1384 struct address_space *mapping = swap_file->f_mapping;
1385 struct inode *inode = mapping->host;
1387 loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
1390 ret = filemap_write_and_wait(inode->i_mapping);
1395 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
1396 ops, &isi, iomap_swapfile_activate_actor);
1404 if (isi.iomap.length) {
1405 ret = iomap_swapfile_add_extent(&isi);
1410 *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
1411 sis->max = isi.nr_pages;
1412 sis->pages = isi.nr_pages - 1;
1413 sis->highest_bit = isi.nr_pages - 1;
1414 return isi.nr_extents;
1416 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
1417 #endif /* CONFIG_SWAP */
1420 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
1421 void *data, struct iomap *iomap)
1423 sector_t *bno = data, addr;
1425 if (iomap->type == IOMAP_MAPPED) {
1426 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
1428 WARN(1, "would truncate bmap result\n");
1435 /* legacy ->bmap interface. 0 is the error return (!) */
1437 iomap_bmap(struct address_space *mapping, sector_t bno,
1438 const struct iomap_ops *ops)
1440 struct inode *inode = mapping->host;
1441 loff_t pos = bno >> inode->i_blkbits;
1442 unsigned blocksize = i_blocksize(inode);
1444 if (filemap_write_and_wait(mapping))
1448 iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
1451 EXPORT_SYMBOL_GPL(iomap_bmap);