Causal mediation analysis is a useful tool to disentangle the total causal effect in order to understand the biological or mechanistic pathways. We consider the setting where both the primary outcome and the mediator are event times. Due to the causal and temporal ordering of the two events, the mediator may be subject to censoring by the occurrence of the primary outcome, but not vice versa, such that we collect semi-competing risks data. Causal mediation analysis with semi-competing risks data has not been studied, even though a number of methods have been proposed for causal mediation analysis with survival outcome. We propose a novel principal stratification approach to identify the causal mediation effects in sub-populations. In particular, we define three principal strata based on the susceptibility of the intermediate event given different treatments and model the marginal structural of the counterfactuals by proportional hazards models. A nonparametric maximum likelihood estimator is proposed and the statistical properties are studied.