The next-generation weak lensing surveys (i.e., LSST, Euclid, and WFIRST) will require exquisite control over systematic effects. In this paper, we address shear calibration and present the most realistic forecast to date for LSST/Euclid/WFIRST and CMB lensing from a stage 4 CMB experiment ("CMB S4"). We use the COSMOLIKE code to simulate a joint analysis of all the two-point functions of galaxy density, galaxy shear, and CMB lensing convergence. We include the full Gaussian and non-Gaussian covariances and explore the resulting joint likelihood with Monte Carlo Markov chains. We constrain shear calibration biases while simultaneously varying cosmological parameters, galaxy biases, and photometric redshift uncertainties. We find that CMB lensing from CMB S4 enables the calibration of the shear biases down to 0.2%-3% in ten tomographic bins for LSST (below the ∼0.5% requirements in most tomographic bins), down to 0.4%-2.4% in ten bins for Euclid, and 0.6%-3.2% in ten bins for WFIRST. For a given lensing survey, the method works best at high redshift where shear calibration is otherwise most challenging. This self-calibration is robust to Gaussian photometric redshift uncertainties and to a reasonable level of intrinsic alignment. It is also robust to changes in the beam and the effectiveness of the component separation of the CMB experiment, and slowly dependent on its depth, making it possible with third-generation CMB experiments such as AdvACT and SPT-3G, as well as the Simons Observatory.