Channel sate acquisition in frequency-division du-plexing (FDD) massive MIMO system is challenging due to the huge feedback overhead. Machine learning (ML) has emerged as a powerful technology to address this challenge. In this paper, we resort to information bottleneck (IB) theoretical principle to design a joint feedback compression, quantization and channel learning algorithm in FDD massive MIMO systems, called IBNet. Compared to the existing ML-based designs, the proposed IBNet can systematically seek for the optimal balance between the channel estimation accuracy and feedback overhead. To auto-matically learn the feedback compression, a sparsity inducing prior is utilized to sparsify the feature vector, thereby reducing the feedback overhead significantly. Furthermore, to improve the generality of proposed IBNet, we propose an adaptive IBNet, which can adapt to different channel conditions with one neural network. Simulation results show that the proposed scheme significantly reduces the feedback overhead, meanwhile improving the channel estimation accuracy.