Knowledge graph (KG) deeply integrates users, items, attributes, and various relationships, which provides a structured contextual environment for explainable RS. However, existing KG-based models can only provide explainability according to feature similarity of users and items, and cannot provide process-based explainability. To address this problem, this paper introduces an explainable recommendation model using global preference paths on KG. We define preference paths that are the relational paths between users and potential items in KG to clearly model the purchasing behaviors of users. Then, we present a reinforcement learning and global attention network (RLGAN) framework to extract preference paths from KG. The framework leverages reinforcement learning agent to traverse complex relational paths, and combines a global attention network to extract diverse preference paths. According to these paths, recommendation items and process-based reasons can be generated. Experimental results on public datasets show that the proposed model not only can achieve much higher performance than the SOTA baseline models, but also can provide enhanced reasons with high credibility on recommendation items.