diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 51a119ac91cd5aad29e2e239f6e35d06d168701d..676cce61cad9d4e01c253b45687a9d1d51e6a4a7 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -3145,8 +3145,22 @@ static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
 		end = ordered_extent->file_offset + ordered_extent->len - 1;
 		clear_extent_uptodate(io_tree, start, end, NULL);
 
-		/* Drop the cache for the part of the extent we didn't write. */
-		btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
+		/*
+		 * Drop extent maps for the part of the extent we didn't write.
+		 *
+		 * We have an exception here for the free_space_inode, this is
+		 * because when we do btrfs_get_extent() on the free space inode
+		 * we will search the commit root.  If this is a new block group
+		 * we won't find anything, and we will trip over the assert in
+		 * writepage where we do ASSERT(em->block_start !=
+		 * EXTENT_MAP_HOLE).
+		 *
+		 * Theoretically we could also skip this for any NOCOW extent as
+		 * we don't mess with the extent map tree in the NOCOW case, but
+		 * for now simply skip this if we are the free space inode.
+		 */
+		if (!btrfs_is_free_space_inode(BTRFS_I(inode)))
+			btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
 
 		/*
 		 * If the ordered extent had an IOERR or something else went