With the large-scale deployment of RDMA in data centers, distributed applications can efficiently transfer data with very low CPU overhead. However, PFC that ensure RDMA lossless environment brings problems such as head-of-line blocking and pause frame storm, which seriously affects network performance. Lossy RDMA is increasingly a focus, but the timeout problem caused by incast packet loss is still the main cause of high tail delay. In this paper, we propose SPLR, a selective packet loss recovery support for RDMA, which can selectively drop packets that will not result in timeout by anticipating the switch overflow in advance. SPLR comprehensively considers traffic characteristics, congestion mitigation and drop cost, establishing a multi-objective optimization model on the congested port to actively drop the selected packets in flow unit, thus, to achieve lower flow completion time and better performance. Experiments show that SPLR achieves up to 8.64 \(\times \) reduction of 99.9th percentile flow completion time compared to IRN, and other alternatives.