There have been growing interests in building a conversational recommender system, where the system simultaneously interacts with the user and explores the user's preference throughout conversational interactions. Recommendation and conversation were usually treated as two separate modules with limited information exchange in existing works, which hinders the capability of both systems: (1) dialog merely incorporated recommendation entities without being guided by an explicit recommendation-oriented policy; (2) recommendation utilized dialog only as a form of interaction instead of improving recommendation effectively. To address the above issues, we propose a novel recommender dialog model: CR-Walker. In order to view the two separate systems within a unified framework, we seek high-level mapping between hierarchical dialog acts and multi-hop knowledge graph reasoning. The model walks on a large-scale knowledge graph to form a reasoning tree at each turn, then mapped to dialog acts to guide response generation. With such a mapping mechanism as a bridge between recommendation and conversation, our framework maximizes the mutual benefit between two systems: dialog as an enhancement to recommendation quality and explainability, recommendation as a goal and enrichment to dialog semantics. Quantitative evaluation shows that our model excels in conversation informativeness and recommendation effectiveness, at the same time explainable on the policy level.