Elderly fall prevention and detection is extremely crucial especially with the fast aging society. In this paper, we propose mmFall, a novel system for fall detection. In terms of sensing modality, we adopt the emerging millimeter-wave (mmWave) radar sensor to collect the point cloud of a moving human body along with its estimated centroid. In terms of detection methodology, we propose a variational recurrent autoencoder (VRAE) model to monitor the motion anomaly level, and claim a fall detection when two conditions meet simultaneously, viz. (i) a anomaly level spike and (ii) a sudden drop of body's centroid height. The mmWave radar sensor provides advantages, such as privacy-compliance and high sensitivity to motion, over the traditional sensing modalities. To overcome the difficulties in machine learning using radar due to the data randomness, the VRAE uses variational inference, a probabilistic approach rather than deterministic approach, to infer the posterior probability of latent motion state at every frame, followed by a recurrent neural network (RNN) to learn the temporal features of the motion over multiple frames. Moreover, to circumvent the difficulties in fall data collection and labeling in the traditional fall detection research, the VRAE is built upon an autoencoder architecture, a semisupervised approach, and trained with only normal activities of daily living (ADL) such that in the inference stage the VRAE will generate a spike in the anomaly level once an abnormal motion, such as fall, occurs. During the experiment, we implemented the VRAE along with two other baselines, and tested on the dataset collected in an apartment. The receiver operating characteristic (ROC) curve indicates that our proposed model outperforms the other two baselines, and achieves 98% detection out of 50 falls at expense of just 2 false alarms.