The channel state information (CSI) plays a pivotal role in realizing precoding design and signal detection for multiple-input multiple-output (MIMO) systems. However, a large number of antennas in massive MIMO systems leads to a huge CSI matrix and impractical feedback overhead. To address this challenge, we propose a novel and efficient CSI feedback network termed Global-Local Feature Reconstruction CsiNet (GLCsiNet), where the network achieves multi-feature extraction of CSI by leveraging global and local feature extraction networks. In contrast to existing deep learning based methods, GLCsiNet integrates the advantageous aspects of recurrent neural networks and convolutional neural networks to more effectively exploit the global and local features of the CSI matrix. Simulation results demonstrate that the proposed GLCsiNet offers notable performance improvements with minimal computational complexity compared to the state-of-the-art method.