Recent advances in radiology report generation (RRG) have shown promising improvements in natural language generation (NLG) metrics. However, clinical efficacy (CE) is often overlooked, leading to diagnostic hallucinations in the generated reports. Moreover, most existing approaches are limited to chest image reports due to their reliance on chest-specific information extraction tools or labels. In this paper, we introduce DiagDiff, a Diagnosis-aware Diffusion model for RRG to improve clinical efficacy. Our framework consists of two stages: diagnosis-aware prompt generation and report generation. First, DiagDiff is designed to generate diagnosis-aware prompts with various pre-trained labelers, allowing it to work beyond chest reports. Second, DiagDiff generates reports through a diffusion model guided by these diagnosis-aware prompts. To mitigate diagnostic hallucinations, we introduce majority voting to select the optimal report from the candidates by weighing NLG and CE metrics. Experimental results demonstrate that DiagDiff significantly outperforms state-of-the-art methods in RRG across four diverse organ image datasets, particularly on clinical efficacy metrics.