We consider a stochastic inventory system with general piecewise linear ordering cost. The cumulative demand is modeled as a Brownian motion process. The ordering cost function is neither convex nor concave; it may not be monotone; and it is not even necessarily continuous, and it includes most ordering cost functions studied in the literature, e.g., economies-of-scale or dis-economies of scale, all-unit discount or incremental discount, and multiple setup costs, as special cases. In addition to ordering cost, the system incurs the usual holding/shortage cost, and the objective is to minimize the average system cost per unit of time. Despite the complexity in the ordering cost function, we show that an optimal control policy is very simple: it is either an (s, S) policy or a one-sided singular control policy.