In data center hybrid traffic (TCP/RDMA) scenarios, traditional dynamic buffer management schemes suffer from unfair buffer resource contention due to the high Round-Trip Time characteristics and aggressive TCP traffic packet retransmission mechanisms. This frequent contention triggers Priority Flow Control pause frames, causing tail latency spikes. Existing methods primarily focus on local buffer management in switches, lacking cross-layer collaboration with higher-layer transport protocols. This results in slow network convergence and difficulty in achieving global resource optimization. This paper proposes Coord-BM, a cross-layer collaborative buffer management framework addressing buffer contention in hybrid TCP/RDMA data center networks. Coord-BM introduces dynamic threshold feedback to host-side congestion control, coupled with a deep reinforcement learning-based online optimizer that jointly tunes throughput, latency, and fairness through weight-adaptive multi-objective utility. An extended congestion notification packet or acknowledge character field enables low-overhead switch-host state synchronization. Evaluations demonstrate Coord-BM reduces PFC pause frames by 62.5%, slashes RDMA tail latency by 64.2%, and improves hybrid traffic buffer utilization by 31.2% compared to the dynamic threshold method, achieving holistic end-to-end resource optimization.