Graph Convolution Network (GCN) is a potent deep learning methodology and GCN-based graph data augmentation methods excel in recommender systems. The augmented data, as the higher-order collaborative signals, is generated by pre-training the GCN model on raw data, before making the final prediction. Despite their success, three problems have always existed. First, during the pre-training phase, over-smooth and data sparsity have a serious impact on the quality of augmented graph data. Second, the influence degree of higher-order collaborative signals are difficult to control, resulting in unstable embeddings. Last, the output of the pre-trained model is only used to augment the data and is not utilized in subsequent training phases, leading to performance penalty. In this study, we propose an efficient Collaborative Signals Augmentation model based on GCN for Recommendation (CSA4Rec). In the pre-training phase, graph contrastive learning is utilized to generate distinct embeddings to address over-smooth and data sparsity. To manipulate the higher-order collaborative signals effectively, the signals are generated from raw data in the modeling phase and aggregated independently by GCN in the training phase. The Residual-like connection is employed to leverage the output of the pre-trained model to improve the convergence speed in the next training phase. Extensive experiments demonstrate the effectiveness and rationality of CSA4Rec.