We report radio observations of the bright, compact supernova remnant 3C 391 using the Very Large Array of the National Radio Astronomy Observatory (NRAO) at 330, 1468, and 4848 MHz. We present and discuss high-resolution images of total intensity, polarization, and spectral index. The large-scale morphology consists of a bright partial shell with a considerably larger plateau of fainter emission extending past the open end of the shell, suggesting that the remnant is directly interacting with a dense region of the interstellar medium, possibly a molecular cloud. The partial shell may result from gradients in the external magnetic field that would be expected in the presence of strong density gradients. Small-scale extensions beyond the shell edge can be interpreted as due to relativistic electrons diffusing upstream of the shock along external magnetic field lines with a mean free path about an order of magnitude smaller than characteristic of the mean interstellar medium for particles of energy a few GeV. If this interpretation is correct, shock-accelerated electrons are dominantly produced where the shock normal is perpendicular to the upstream magnetic field. We find no polarization at 330 or 1468 MHz, with 3 sigma upper limits of 6 mJy/beam (10 sec beam) at 330 MHz and 0.3 mJy/beam (6 sec beam) at 1468 MHz. We do obseve polarized flux at 4848 MHz, but a mean polarized fraction of 0.77% +/- 0.06 %, far lower than typical for bright supernova remnants. Tangled or disordered magnetic fields in the emitting region of the radio shell may be responsible for depolarizing the radio synchrotron radiation, but some internal Faraday depolarization may also occur. We crudely estimate the foreground Faraday rotation to be about -- 500 rad/sq m, consistent with previous estimates. Spectral index images created from the total intensity images show no variation beyond delta(alpha) = 0.1. We do see variations at lower levels which are formally marginally significant but which are not consistent among the three pairs of frequencies. Small errors in the total flux density at each frequency or in the deconvolutions are probably responsible.