On-Policy Fine-grained Knowledge Feedback for Hallucination Mitigation