Significant discrepancies between Monte Carlo dose calculations and measurements for the Varian 18 MV photon beam with a large field size (40 x 40 cm2) were reported by different investigators. In this work, we investigated these discrepancies based on a new geometry model ("New Model") of the Varian 21EX linac using the GEPTS Monte Carlo code. Some geometric parameters used in previous investigations (Old Model) were inaccurate, as suggested by Chibani in his AAPM presentation (2004) and later confirmed by the manufacturer. The entrance and exit radii of the primary collimator of the New Model are 2 mm larger than previously thought. In addition to the corrected dimensions of the primary collimator, the New Model includes approximate models for the lead shield and the mirror frame between the monitor chamber and the Y jaws. A detailed analysis of the phase space data shows the effects of these corrections on the beam characteristics. The individual contributions from the linac component to the photon and electron fluences are calculated. The main source of discrepancy between measurements and calculations based on the Old Model is the underestimated electron contamination. The photon and electron fluences at the isocenter are 5.3% and 36% larger in the New Model in comparison with the Old Model. The flattening filter and the lead shield (plus the mirror frame) contribute 48.7% and 13% of the total electron contamination at the isocenter, respectively. For both open and filtered (2 mm Pb) fields, the calculated (New Model) and measured dose distributions are within 1% for depths larger than 1 cm. To solve the residual problem of large differences at shallow depths (8% at 0.25 cm depth), the detailed geometry of an IC-10 ionization chamber was simulated and the dose in the air cavity was calculated for different positions on the central axis including at the surface, where half of the chamber is outside the phantom. The calculated and measured chamber responses are within 3% even at the zero depth.