The traditional adaptive modulation and coding (AMC) scheme has been designed separately from higher protocol layers. In this paper, we consider the joint effects of finite-length buffer at the data link layer and the channel state information (CSI) at the physical layer for the adaptive transmission over the multiple input multiple output (MIMO) fading channel. A novel cross-layer adaptive transmission policy is proposed to minimize the system overall packet loss rate (PLR) including the packet dropping rate and the packet error rate. We model the policy decision-making process as a Markov decision problem and utilize linear programming method to obtain the optimal adaptive transmission policy. Simulation results demonstrate the verification of the performance analysis, and show the great improvement of the system PLR compared with the traditional AMC scheme.