> +static int blk_iolatency_try_init(char *input)
> +{
> +	static DEFINE_MUTEX(init_mutex);
> +	struct block_device *bdev;
> +	int ret = 0;
> +
> +	bdev = blkcg_conf_open_bdev(&input);
> +	if (IS_ERR(bdev))
> +		return PTR_ERR(bdev);

> +retry:
>  	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
> +	if (ret == -EOPNOTSUPP) {
> +		ret = blk_iolatency_try_init(buf);

It's a little sad to do two block device lookups here (even if it
obviously doesn't matter for performance).  I wonder if it would
make sense to explicitly support the lazy init pattern
in blkg_conf_prep somehow.

Otherwise I'm all for the lazy init.