We propose a comprehensive scheme for realizing a massive multiple-input multiple-output (MIMO) system with dual-polarized antennas in frequency division duplexing (FDD) mode. Dual-polarized arrays are commonly employed due to the favorable property that, in principle, they can double the number of channel spatial degrees of freedom with a less-than-proportional increase in array size. However, processing a dual-polarized massive MIMO channel is demanding due to the high channel dimension and the lack of Uplink-Downlink (UL-DL) channel reciprocity in FDD mode. In particular, the difficulty arises in common channel training and DL precoding in a multi-user setup. To address this, we develop a unified framework consisting of three steps: (1) covariance estimation to efficiently estimate the UL covariance from noisy UL pilots; (2) a UL-DL covariance transformation method that obtains the DL covariance from the estimated UL covariance, eliminating the need for DL channel covariance training via pilot transmission; <xref ref-type="disp-formula" rid="deqn3a-deqn3b" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">(3)</xref> a joint multi-user DL channel training method, which enables the BS to estimate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">effective DL channels</i> given any protocol-specific pilot dimension and to use them for interference-free DL beamforming and data transmission. Through extensive simulations, we show that our scheme is applicable to a variety of communication scenarios in terms of the number of antennas, UL and DL pilot dimensions, and angular scattering properties. Unlike the common trend in the literature, we do <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">not</i> make strong structural assumptions about the wireless channel (such as angular sparsity), ensuring a general treatment of the problem.