We use hydrodynamical simulations from the OverWhelmingly Large Simulations (OWLS) project to investigate the dependence of the physical properties of galaxy populations at redshift 2 on metal-line cooling and feedback from star formation and active galactic nuclei (AGN). We find that if the sub-grid feedback from star formation is implemented kinetically, the feedback is only efficient if the initial wind velocity exceeds a critical value. This critical velocity increases with galaxy mass and also if metal-line cooling is included. This suggests that radiative losses quench the winds if their initial velocity is too low. If the feedback is efficient, then the star formation rate is inversely proportional to the amount of energy injected per unit stellar mass formed (which is proportional to the initial mass loading for a fixed wind velocity). This can be understood if the star formation is self-regulating, i.e. if the star formation rate (and thus the gas fraction) increases until the outflow rate balances the inflow rate. Feedback from AGN is efficient at high masses, while increasing the initial wind velocity with gas pressure or halo mass allows one to generate galaxy-wide outflows at all masses. Matching the observed galaxy mass function requires efficient feedback. In particular, the predicted faint-end slope is too steep unless we resort to highly mass loaded winds for low-mass objects. Such efficient feedback from low-mass galaxies (M* ≪ 1010 M⊙) also reduces the discrepancy with the observed specific star formation rates, which are higher than predicted unless the feedback transitions from highly efficient to inefficient just below M* ˜ 5 × 109 M⊙.