Trajectory similarity computation is a fundamental task in various spatial information processing applications. Traditional similarity metrics typically treat trajectories as ordinary sequential data and employ deep learning models designed with multi-layer LSTM backends. However, these methods suffer from two critical limitations: (i) inadequate utilization of important spatiotemporal features inherent in trajectory data, such as travel semantics; (ii) inefficient global modeling of long trajectories. To address these issues, we propose a novel Trajectory similarity computation based on a Bidirectional State Space Model, namely TrajBISSM. Specifically, we designed a dynamic knowledge graph strategy and a trajectory encoder based on the state space model. In particular, the dynamic knowledge graph strategy innovatively introduces the dynamic transition probability matrix of the road network to affect the node embedding, enabling the point embedding to contain richer spatiotemporal feature information. The trajectory encoder achieves the global semantic modeling of long trajectories through independent forward and backward state propagation. Experiments on two real-world datasets show that TrajBISSM significantly outperforms the existing methods in trajectory similarity computation.