Accurate and efficient root cause localization is essential for ensuring the reliability of cloud-native systems. However, most existing approaches focus on identifying the root cause of failures within single-level resource entities, such as servers, microservices or pods, while overlooking the fault propagation across entities at different levels. Moreover, many methods struggle to adapt to the dynamic nature of these entities, making efficient root cause localization a significant challenge. In this paper, we propose CFRCL, a cross-level failure root cause localization framework built from a holistic fault propagation perspective. CFRCL first detects anomalies to identify fault-related entities, and then integrates these results with microservice deployment data to generate a causal skeleton, narrowing the search space. It constructs a causal graph from abnormal metrics within the skeleton and introduces a novel initialization method to assign node scores. A transition probability matrix is then built, and a random walk algorithm simulates the fault propagation to rank the potential root causes. Experiments on two benchmark datasets demonstrate that CFRCL outperforms existing methods, achieving state-of-the-art localization accuracy.