The unpredictable activities of primary users (PUs) make the channel availabilities in cognitive radio networks (CRNs) very unstable, which causes routing in CRNs to be more difficult than in traditional wireless networks. Specifically, when a source node needs to select a route to reach the destination, the “optimal” route during the route selection phase may not be optimal in the data transmission phase. In this paper, we propose a novel routing protocol based on the traditional source routing protocols. We consider the angle dimension by assuming that directional antennas are equipped on every node, which facilitate the marking of boundaries of PUs. We use the USRP/Gnuradio testbed to show the sensing result differences of different directions at the boundary area of a PU. For every optional route between a source and a destination node, we evaluate its reliability and other performance by evaluating the PU areas through which it passes and by estimating the possible transmission rate. Based on these parameters, we propose an algorithm for route selection, considering both reliability and delay. Our routing protocol only requires very limited piggyback information, and it is highly adaptable under the dynamic channel availabilities. We evaluate our approach through extensive simulations.