Standard methods for estimating average causal effects require complete observations of the exposure and confounders. In observational studies, however, missing data are ubiquitous. We propose methods for estimating average causal effects when exposures and potential confounders may be missing. We consider missingness at random (MAR) and missingness not at random (MNAR) assumptions. Under each setting, we show that the average causal effects are non-parametrically identified and propose targeted maximum likelihood estimators that are semi-parametric efficient and doubly robust, allowing misspecification of either (i) the outcome models or (ii) the exposure and missingness models. The proposed methods are suitable for binary (or any other) outcome types, and we apply them to a motivating study of the effect of opioid usage on all-cause mortality in the National Health and Nutrition Examination Survey (NHANES) data.