The existing methods based on convex-optimization theory, which use the concept of SINR, can just design the optimal precoder for each user with single antenna. In this paper, we design the optimal precoding matrices for multi-user MIMO downlinks by solving the optimization problem that minimizes total transmit power subject to signal-leakage-plus-noise-ratio (SLNR) constraints. Because SLNR of each user is determined by its own precoding matrix and is independent of other users, the goal problem can be separated into a series of decoupled low-complexity quadratically constrained quadratic programs (QCQPs). Using the semidefinite relaxation (SDR) technique, these QCQPs can be reformulated into the semidefinite programs (SDP) and be solved effectively. Simulation results show that proposed precoding scheme is quite feasible when each user has two receive antennas, and it has better bit error rate (BER) performance than the original maximal-SLNR precoding scheme when SLNR of each user satisfies large threshold value.