In Cognitive Radio (CR) networks, there could be a spectrum market which operates in real time with primary users (PU) as manager, where the secondary users (SU) pay the PUs for spectrum resource usage. Multi-carrier systems such as OFDM are best candidates for applying in CR networks because of the spectrum shaping and high adaptive capabilities. In this paper, an optimal resource allocation scheme aims at maximizing PU's reward in OFDM-based CR networks is proposed. Both the interference limit and BER requirements of SUs are considered by the PU to allocate its spectrum resources. The scheme is modeled as restless bandits problem, which can dramatically simplify the computation and implementation. Furthermore, extensive simulation results show that our proposed scheme can improve the reward of PU significantly compared to the existing random scheme and greedy scheme.