1 files changed, 29 insertions, 13 deletions

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 0067de1b38fc..06137cf281bc 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -926,7 +926,7 @@ static noinline int cow_file_range(struct inode *inode,
 	u64 alloc_hint = 0;
 	u64 num_bytes;
 	unsigned long ram_size;
-	u64 disk_num_bytes;
+	u64 min_alloc_size;
 	u64 cur_alloc_size;
 	u64 blocksize = root->sectorsize;
 	struct btrfs_key ins;
@@ -942,7 +942,6 @@ static noinline int cow_file_range(struct inode *inode,
 
 	num_bytes = ALIGN(end - start + 1, blocksize);
 	num_bytes = max(blocksize,  num_bytes);
-	disk_num_bytes = num_bytes;
 
 	/* if this is a small write inside eof, kick off defrag */
 	if (num_bytes < 64 * 1024 &&
@@ -969,18 +968,33 @@ static noinline int cow_file_range(struct inode *inode,
 		}
 	}
 
-	BUG_ON(disk_num_bytes >
-	       btrfs_super_total_bytes(root->fs_info->super_copy));
+	BUG_ON(num_bytes > btrfs_super_total_bytes(root->fs_info->super_copy));
 
 	alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 	btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 
-	while (disk_num_bytes > 0) {
+	/*
+	 * Relocation relies on the relocated extents to have exactly the same
+	 * size as the original extents. Normally writeback for relocation data
+	 * extents follows a NOCOW path because relocation preallocates the
+	 * extents. However, due to an operation such as scrub turning a block
+	 * group to RO mode, it may fallback to COW mode, so we must make sure
+	 * an extent allocated during COW has exactly the requested size and can
+	 * not be split into smaller extents, otherwise relocation breaks and
+	 * fails during the stage where it updates the bytenr of file extent
+	 * items.
+	 */
+	if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
+		min_alloc_size = num_bytes;
+	else
+		min_alloc_size = root->sectorsize;
+
+	while (num_bytes > 0) {
 		unsigned long op;
 
-		cur_alloc_size = disk_num_bytes;
+		cur_alloc_size = num_bytes;
 		ret = btrfs_reserve_extent(root, cur_alloc_size,
-					   root->sectorsize, 0, alloc_hint,
+					   min_alloc_size, 0, alloc_hint,
 					   &ins, 1, 1);
 		if (ret < 0)
 			goto out_unlock;
@@ -1033,7 +1047,7 @@ static noinline int cow_file_range(struct inode *inode,
 				goto out_drop_extent_cache;
 		}
 
-		if (disk_num_bytes < cur_alloc_size)
+		if (num_bytes < cur_alloc_size)
 			break;
 
 		/* we're not doing compressed IO, don't unlock the first
@@ -1050,8 +1064,10 @@ static noinline int cow_file_range(struct inode *inode,
 					     start + ram_size - 1, locked_page,
 					     EXTENT_LOCKED | EXTENT_DELALLOC,
 					     op);
-		disk_num_bytes -= cur_alloc_size;
-		num_bytes -= cur_alloc_size;
+		if (num_bytes < cur_alloc_size)
+			num_bytes = 0;
+		else
+			num_bytes -= cur_alloc_size;
 		alloc_hint = ins.objectid + ins.offset;
 		start += cur_alloc_size;
 	}
@@ -8301,7 +8317,6 @@ static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
 	bio->bi_private = dip;
 	bio->bi_end_io = btrfs_end_dio_bio;
 	btrfs_io_bio(bio)->logical = file_offset;
-	atomic_inc(&dip->pending_bios);
 
 	while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
 		if (map_length < submit_len + bvec->bv_len ||
@@ -8358,7 +8373,8 @@ submit:
 	if (!ret)
 		return 0;
 
-	bio_put(bio);
+	if (bio != orig_bio)
+		bio_put(bio);
 out_err:
 	dip->errors = 1;
 	/*
@@ -8405,7 +8421,7 @@ static void btrfs_submit_direct(int rw, struct bio *dio_bio,
 	io_bio->bi_private = dip;
 	dip->orig_bio = io_bio;
 	dip->dio_bio = dio_bio;
-	atomic_set(&dip->pending_bios, 0);
+	atomic_set(&dip->pending_bios, 1);
 	btrfs_bio = btrfs_io_bio(io_bio);
 	btrfs_bio->logical = file_offset;