The improvements of random walk search mainly depend on allocating weight for neighbor peers, which is always incurred on a high overhead and are not very helpful for rare items. This paper proposes a bidirectional random walk search mechanism (short for BRWS) for unstructured P2P network, according to the analysis of basic properties about random walk as well as the special property that random walk tends toward high degree nodes. The mechanism is proved theoretically in this paper, and can improve search success rate, including searching for rare items. It also has a high tolerance for churn. In the static and dynamic environment, comparisons were made among Random Walk, APS (adaptive probabilistic search), PQR (path-traceable query routing), P2PBSN (peer-to- peer based on social network) and BRWS based on three topologies: Random graph, scale free, small world. The experimental results show that BRWS can actually improve the search success rate with lower overhead even when searching rare resources. The method proposed in this paper can apply in P2P file sharing networks.