Single image super-resolution (SISR) has achieved great progress based on convolutional neural networks (CNNs) such as generative adversarial network (GAN). However, most deep learning architectures cannot utilize the hierarchical features in original low-resolution images, which may result in the loss of image details. To recover visually high-quality high-resolution images, we propose a novel Multi-recursive residual dense Attention Generative Adversarial Network (MAGAN). Our MAGAN enjoys the ability to learn more texture details and overcome the weakness of conventional GAN-based models, which easily generate redundant information. In particular, we design a new multi-recursive residual dense network as a module in our generator to take advantage of the information from hierarchical features. We also introduce a multi-attention mechanism to our MAGAN to capture more informative features. Moreover, we present a new convolutional block in our discriminator by utilizing switchable normalization and spectral normalization to stabilize the training and accelerate convergence. Experimental results on benchmark datasets indicate that MAGAN yields finer texture details and does not produce redundant information in comparison with existing methods.