The correlative and uncertain relationship between software metrics and software reliability was analyzed. An approach for software reliability qualitative evaluation by using Bayesian networks was proposed. The rationale and process of this method were described in detail. The software reliability evaluation results were obtained by the causality reasoning. The proposed approach has been implemented in a case study. The experimental results constructed by the application process, key step and computational solution, showed that the approach was effective and could combine the expert experience, history data and various kinds of uncertain information to improve the modeling efficiency and validity, and provided a new method for software reliability evaluation.