Missing value imputation is a fundamental challenge in spatio-temporal data modeling. To address this issue, we propose a novel model named Bidirectional Mamba Imputation (BMI), which incorporates the selective state space model, Mamba, into the missing data recovery task. The proposed BMI leverages bidirectional information flow to capture both past and future dependencies in time series, and introduces a spatio-temporal block design to jointly learn temporal and spatial patterns. Moreover, a selective mechanism is embedded in the state transition process to adaptively focus on crucial variables and time steps. Extensive experiments conducted on four real-world datasets demonstrate that BMI significantly outperforms existing traditional and deep learning baselines under both point and block missing scenarios. Ablation studies confirm the effectiveness of the bidirectional design and the selective state space module. Additionally, efficiency analysis shows that BMI is well-suited for deployment in resource-constrained environments due to its high memory and I/O efficiency.