Fall detection is of great importance for elderly care. WiFi-based fall detection has advantages over visions and wearables wrt. privacy protection, convenience, low-cost and ubiquity. However, environment dependence is a major challenge that hinders the real-world deployment of such systems. In this paper, we investigate the problems and present an environment independent fall detection method exploiting WiFi Channel State Information (CSI). To achieve environment independence, we propose a feature disentanglement neural network to separate the motion-related features and the environment-related features. Only the motion-related features are extracted to classify falls and non-falls. To achieve real-time fall detection, we propose an online data segmentation method to detect and extract the motion segments from continuous CSI streams automatically. To mitigate the scarcity of fall data and improve the robustness of the model, we design a composite Autoencoder to generate virtual fall samples by adding random noise to real fall samples. Extensive real-world evaluations show that the proposed method is free of training in new environments and achieves real-time fall detection independent of users, locations, rooms, and times.