This paper addresses the H-2 filtering design problem of discrete-time Markov jump linear systems. First, under the assumption that the Markov parameter is measured, the main contribution is on the LMI characterisation of all filters such that the estimation error remains bounded by a given H-2 norm level, yielding the complete solution of the mode-dependent filtering design problem. Based on this result, a robust filter design to deal with convex bounded parameter uncertainty is considered. Second, from the same LMI characterisation, a design procedure for mode-independent filtering design is proposed. Some examples are solved for illustration and comparison.