Cyber-Physical Systems (CPSs) are combinations of physical entities controlled by software systems to accomplish specified tasks under stringent real-time and physical entity constraints. As more and more physical entities are equipped with embedded computers, they are becoming more and more intelligent. However, the problem of effectively composing the services provided by cyber and physical entities to achieve specific goals still remains a challenge. Traditional service-oriented models and composition techniques are insufficient for CPS. In this paper, we present a novel Physical-Entity (PE) service-oriented model to address the problem, including the concepts of PE-ontology and PE-SOA specifications. Based on the model, we develop a two-level compositional reasoning approach, which divides the process into abstract and physical levels to expedite the composition process. With the assistance of the PE-ontology and PE-SOA, abstract level reasoning is performed by hiding the physical level details. This separation greatly reduces the search space for both levels through a divide-and-conquer technique. The model and the composition approach are illustrated using a simplified emergency response case study system.