[dpdk-stable] [PATCH 18.11] net/mlx5: fix vectorized Rx burst termination

[Alexander Kozyrev]

From: Alexander Kozyrev <akozyrev at mellanox.com>

[ upstream commit c9cc554ba423641d3515671269c5648dab3bb9ad ]

Maximum burst size of Vectorized Rx burst routine is set to
MLX5_VPMD_RX_MAX_BURST(64). This limits the performance of any
application that would like to gather more than 64 packets from
the single Rx burst for batch processing (i.e. VPP).

The situation gets worse with a mix of zipped and unzipped CQEs.
They are processed separately and the Rx burst function returns
small number of packets every call.

Repeat the cycle of gathering packets from the vectorized Rx routine
until a requested number of packets are collected or there are no
more CQEs left to process.

Fixes: 6cb559d67b83 ("net/mlx5: add vectorized Rx/Tx burst for x86")
Cc: stable at dpdk.org

Signed-off-by: Alexander Kozyrev <akozyrev at mellanox.com>
Acked-by: Slava Ovsiienko <viacheslavo at mellanox.com>
Acked-by: Matan Azrad <matan at mellanox.com>
---
 drivers/net/mlx5/mlx5_rxtx_vec.c      | 17 ++++++++++++-----
 drivers/net/mlx5/mlx5_rxtx_vec_neon.h | 13 ++++++++++---
 drivers/net/mlx5/mlx5_rxtx_vec_sse.h  | 13 ++++++++++---
 3 files changed, 32 insertions(+), 11 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxtx_vec.c b/drivers/net/mlx5/mlx5_rxtx_vec.c
index b8d624d14..ca2389b5d 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec.c
+++ b/drivers/net/mlx5/mlx5_rxtx_vec.c
@@ -232,13 +232,20 @@ uint16_t
 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 {
 	struct mlx5_rxq_data *rxq = dpdk_rxq;
-	uint16_t nb_rx;
+	uint16_t nb_rx = 0;
+	uint16_t tn = 0;
 	uint64_t err = 0;
+	bool no_cq = false;
 
-	nb_rx = rxq_burst_v(rxq, pkts, pkts_n, &err);
-	if (unlikely(err))
-		nb_rx = rxq_handle_pending_error(rxq, pkts, nb_rx);
-	return nb_rx;
+	do {
+		nb_rx = rxq_burst_v(rxq, pkts + tn, pkts_n - tn, &err, &no_cq);
+		if (unlikely(err))
+			nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
+		tn += nb_rx;
+		if (unlikely(no_cq))
+			break;
+	} while (tn != pkts_n);
+	return tn;
 }
 
 /**
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
index cdf9c65f1..566b3ea24 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
@@ -641,13 +641,15 @@ rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
  * @param[out] err
  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
  *   packet to handle.
+ * @param[out] no_cq
+ *   Pointer to a boolean. Set true if no new CQE seen.
  *
  * @return
  *   Number of packets received including errors (<= pkts_n).
  */
 static inline uint16_t
 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
-	    uint64_t *err)
+	    uint64_t *err, bool *no_cq)
 {
 	const uint16_t q_n = 1 << rxq->cqe_n;
 	const uint16_t q_mask = q_n - 1;
@@ -754,8 +756,10 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
 	/* Not to cross queue end. */
 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
-	if (!pkts_n)
+	if (!pkts_n) {
+		*no_cq = !rcvd_pkt;
 		return rcvd_pkt;
+	}
 	/* At this point, there shouldn't be any remained packets. */
 	assert(rxq->rq_pi == rxq->cq_ci);
 	/*
@@ -989,8 +993,10 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 			break;
 	}
 	/* If no new CQE seen, return without updating cq_db. */
-	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
+	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
+		*no_cq = true;
 		return rcvd_pkt;
+	}
 	/* Update the consumer indexes for non-compressed CQEs. */
 	assert(nocmp_n <= pkts_n);
 	rxq->cq_ci += nocmp_n;
@@ -1016,6 +1022,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	}
 	rte_cio_wmb();
 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
+	*no_cq = !rcvd_pkt;
 	return rcvd_pkt;
 }
 
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
index fb384efde..b83fd7f9b 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
@@ -643,13 +643,15 @@ rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq, __m128i cqes[4],
  * @param[out] err
  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
  *   packet to handle.
+ * @param[out] no_cq
+ *   Pointer to a boolean. Set true if no new CQE seen.
  *
  * @return
  *   Number of packets received including errors (<= pkts_n).
  */
 static inline uint16_t
 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
-	    uint64_t *err)
+	    uint64_t *err, bool *no_cq)
 {
 	const uint16_t q_n = 1 << rxq->cqe_n;
 	const uint16_t q_mask = q_n - 1;
@@ -737,8 +739,10 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
 	/* Not to cross queue end. */
 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
-	if (!pkts_n)
+	if (!pkts_n) {
+		*no_cq = !rcvd_pkt;
 		return rcvd_pkt;
+	}
 	/* At this point, there shouldn't be any remained packets. */
 	assert(rxq->rq_pi == rxq->cq_ci);
 	/*
@@ -939,8 +943,10 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 			break;
 	}
 	/* If no new CQE seen, return without updating cq_db. */
-	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
+	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
+		*no_cq = true;
 		return rcvd_pkt;
+	}
 	/* Update the consumer indexes for non-compressed CQEs. */
 	assert(nocmp_n <= pkts_n);
 	rxq->cq_ci += nocmp_n;
@@ -966,6 +972,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	}
 	rte_compiler_barrier();
 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
+	*no_cq = !rcvd_pkt;
 	return rcvd_pkt;
 }
 
-- 
2.24.1