We investigate a cognitive radio system where secondary user (SU) relays primary user (PU) packets using two-phase relaying. SU transmits its own packets with some access probability in relaying phase using time sharing. PU and SU have queues of finite capacity which results in packet loss when the queues are full. Utilizing knowledge of relay queue state, SU aims to maximize its packet throughput while keeping packet loss probability of PU below a threshold. By exploiting structure of the problem, we formulate it as a linear program and find optimal access policy of SU. We also propose low complexity sub-optimal access policies, namely constant probability transmission and step transmission. Numerical results are presented to compare performance of proposed methods and study effect of queue sizes on packet throughput.