This paper presents a fully decentralized just-in-time workflow scheduling method in a P2P Grid system. The proposed solution allows each peer node to autonomously dispatch inter-dependent tasks of workflows to run on geographically distributed computers. To reduce the workflow completion time and enhance the overall execution efficiency, not only does each node perform as a scheduler to distribute its tasks to execution nodes (or resource nodes), but the resource nodes will also set the execution priorities for the received tasks. By taking into account the unpredictability of tasks' finish time, we devise an efficient task scheduling heuristic, namely dynamic shortest makespan first (DSMF), which could be applied at both scheduling phases for determining the priority of the workflow tasks. We compare the performance of the proposed algorithm against seven other heuristics by simulation. Our algorithm achieves 20%~60% reduction on the average completion time and 37.5%~90% improvement on the average workflow execution efficiency over other decentralized algorithms.