In recent years, due to the characteristics of graph convolutional networks (GCN) that can effectively integrate node information and topological structure, they have shown strong capabilities in graph data representation learning. Since the data in the social recommender systems exists in a graph structure, these characteristics of GNN greatly promote the development of social recommendation. However, the construction of GCN-based social recommender systems faces the following challenges: (1) Users exist in both user-item interaction graphs and social relationship graph, and items exist in both user-item interaction graphs and item collaborative similarity graphs; (2) There is high-order connectivity of the user-user social relationship and the item-item collaborative similar relationship; (3) The embeddings of users and items come from different two graphs have different semantic contributions. In order to solve the above three challenges simultaneously, this paper proposes G raph C onvolution Network for S ocial R ecommendation (GCSR), which models the user-item interaction network, social network and item collaborative similar network into a unified network. Specifically, we provide a principled method to explicitly capture the high-order connectivity of user-items to improve embedding. In addition, the gating mechanism is used for feature fusion to model heterogeneous strengths. A large number of experiments on two real datasets prove the effectiveness of the proposed framework GCSR.