Enhancement of reverberant speech is a challenging problem, especially for scenarios with multiple simultaneous sources and moving sources. Multi channel linear prediction is often used for late reverberation cancellation, and the prediction residual is taken as the desired signal. Though the method works in practice with multiple sources also, the source selectivity is absent in the formulation. Spatial filters have better spatial selectivity, but the performance depends on the number of microphones, and the reverberation component in the direction of the source can not be suppressed. A cascade of MCLP followed by spatial filtering is also explored in the literature. In this work, we consider joint MCLP filtering and spatial filtering. We consider both stationary and dynamic source scenarios. In the stationary case, we consider batch estimation using a distortion-less response constrained maximum likelihood estimation. The solution involves solving MCLP and spatial filtering problems, and source RTF estimation in each iteration of the algorithm. A linear dynamical system model is considered for the time variation of the MCLP prediction coefficients. The method is combined with online spatial filtering. The spatial filtered signal is used for estimating the desired signal PSD, required for the computation of MCLP filters. Better performance compared to the methods in literature is obtained. Further, the method provides for spatial selectivity, which is used for desired source extraction in multi source scenarios.