btrfs: fix ordered extent split error handling in btrfs_dio_submit_io

When the call to btrfs_extract_ordered_extent in btrfs_dio_submit_io
fails to allocate memory for a new ordered_extent, it calls into the
btrfs_dio_end_io for error handling.  btrfs_dio_end_io then assumes that
bbio->ordered is set because it is supposed to be at this point, except
for this error handling corner case.  Try to not overload the
btrfs_dio_end_io with error handling of a bio in a non-canonical state,
and instead call btrfs_finish_ordered_extent and iomap_dio_bio_end_io
directly for this error case.

Reported-by: syzbot <syzbot+5b82f0e951f8c2bcdb8f@syzkaller.appspotmail.com>
Fixes: b41b6f6 ("btrfs: use btrfs_finish_ordered_extent to complete direct writes")
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Tested-by: syzbot <syzbot+5b82f0e951f8c2bcdb8f@syzkaller.appspotmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/inode.c

@@ -7873,8 +7873,11 @@ static void btrfs_dio_submit_io(const struct iomap_iter *iter, struct bio *bio,
 
 		ret = btrfs_extract_ordered_extent(bbio, dio_data->ordered);
 		if (ret) {
-			bbio->bio.bi_status = errno_to_blk_status(ret);
-			btrfs_dio_end_io(bbio);
+			btrfs_finish_ordered_extent(dio_data->ordered, NULL,
+						    file_offset, dip->bytes,
+						    !ret);
+			bio->bi_status = errno_to_blk_status(ret);
+			iomap_dio_bio_end_io(bio);
 			return;
 		}
 	}