The resource allocation problem on the downlink of a multiuser OFDM-based cognitive radio (CR) system is formulated using a cross-layer (MAC and PHY layers) approach with the aim of satisfying quality of service (QoS) requirements in real-time applications. The number of subchannels available to the CR system is time-varying as a result of the stochastic nature of the activities of the primary users (PUs). The MAC layer QoS requirements are dynamically converted to PHY layer rate requirements; the conversion depends on the delivery status of queued packets as well as the number of available subchannels. Simulation results show that the proposed cross-layer resource allocation algorithm can provide substantial transmit power reductions compared to existing PHY layer and MAC layer solutions designed for multiuser OFDM systems.