The ai-code-review check on this patch (https://dpdk.org/patch/173059) says that in the enhanced layout cq_ci should move by one entry per compressed CQE rather than by MLX5_CQE_NUM_MINIS() + 1, since all mini-CQEs of an array sit in one 64-byte entry.
I do not think so. In the enhanced layout an array of N mini-CQEs is written into one entry, but it still takes N entries of the CQ: the NIC leaves the other N - 1 reserved (02a6195cbeaa: "The invalidation of reserved CQEs between miniCQEs arrays is not needed"), and the next CQE is N entries further. The Rx burst advances the consumer index the same way: - mlx5_rx_poll_len(): zip->cqe_cnt = MLX5_CQE_NUM_MINIS(op_own) + 1, zip->cq_ci = rxq->cq_ci + zip->cqe_cnt, and rxq->cq_ci = zip->cq_ci once the array is consumed. - rxq_cq_decompress_v(): the next array is looked up at &cq[mcqe_n] with check_cqe_iteration(cq, rxq->cqe_n, rxq->cq_ci + pkts_n), and rxq_burst_v() then does rxq->cq_ci += rxq->decompressed. - The kernel driver does the same in mlx5e_decompress_enhanced_cqe(): wq->cc moves by get_cqe_enhanced_num_mini_cqes() per array. So in this layout, as in the basic one, the number of packets of a compressed CQE and the number of CQ entries it takes are the same value, and rx_queue_count() moves cq_ci by it. Moving by one would land on a reserved entry the NIC did not write in this pass. With this patch, on ConnectX-6 Dx with rxq_cqe_comp_en=129 at 100G, the reported depth stayed within 0..568 of a 32768-entry ring, and the datapath threads that poll it no longer hang. The ternary in the loop condition is the same one mlx5_rx_poll_len() uses; I can move it into a helper in a v2 if you prefer. Thanks, Daniil Bobryshev ([email protected])

