From: Yongseok Koh Date: Tue, 17 Jan 2017 02:09:40 +0000 (-0800) Subject: net/mlx5: add prefetching Rx completion queue X-Git-Tag: spdx-start~4623 X-Git-Url: http://git.droids-corp.org/?a=commitdiff_plain;h=d2e842d01e861bf489be4d34ab63ca17d50ae4e5;p=dpdk.git net/mlx5: add prefetching Rx completion queue On receiving a compressed session of Rx completion, prefetch every entries to be invalidated. Also, invalidate consumed completions per every 8 mini-completions, not to wait until the last entry is consumed. This helps to reduce jitter in rx_burst. Signed-off-by: Yongseok Koh Acked-by: Adrien Mazarguil --- diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index 56c7f780fb..6c23899a4c 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -1157,6 +1157,7 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe, struct rxq_zip *zip = &rxq->zip; uint16_t cqe_n = cqe_cnt + 1; int len = 0; + uint16_t idx, end; /* Process compressed data in the CQE and mini arrays. */ if (zip->ai) { @@ -1167,6 +1168,14 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe, len = ntohl((*mc)[zip->ai & 7].byte_cnt); *rss_hash = ntohl((*mc)[zip->ai & 7].rx_hash_result); if ((++zip->ai & 7) == 0) { + /* Invalidate consumed CQEs */ + idx = zip->ca; + end = zip->na; + while (idx != end) { + (*rxq->cqes)[idx & cqe_cnt].op_own = + MLX5_CQE_INVALIDATE; + ++idx; + } /* * Increment consumer index to skip the number of * CQEs consumed. Hardware leaves holes in the CQ @@ -1176,8 +1185,9 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe, zip->na += 8; } if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) { - uint16_t idx = rxq->cq_ci + 1; - uint16_t end = zip->cq_ci; + /* Invalidate the rest */ + idx = zip->ca; + end = zip->cq_ci; while (idx != end) { (*rxq->cqes)[idx & cqe_cnt].op_own = @@ -1213,7 +1223,7 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe, * special case the second one is located 7 CQEs after * the initial CQE instead of 8 for subsequent ones. */ - zip->ca = rxq->cq_ci & cqe_cnt; + zip->ca = rxq->cq_ci; zip->na = zip->ca + 7; /* Compute the next non compressed CQE. */ --rxq->cq_ci; @@ -1222,6 +1232,13 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe, len = ntohl((*mc)[0].byte_cnt); *rss_hash = ntohl((*mc)[0].rx_hash_result); zip->ai = 1; + /* Prefetch all the entries to be invalidated */ + idx = zip->ca; + end = zip->cq_ci; + while (idx != end) { + rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]); + ++idx; + } } else { len = ntohl(cqe->byte_cnt); *rss_hash = ntohl(cqe->rx_hash_res);