[dpdk-dev] [PATCH v5 16/25] mlx5: replace countdown with threshold for Tx completions

[Nelio Laranjeiro <nelio.laranjeiro@6wind.com>]

From: Adrien Mazarguil <adrien.mazarguil@6wind.com>

Replacing the variable countdown (which depends on the number of
descriptors) with a fixed relative threshold known at compile time improves
performance by reducing the TX queue structure footprint and the amount of
code to manage completions during a burst.

Completions are now requested at most once per burst after threshold is
reached.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
---
 drivers/net/mlx5/mlx5_defs.h |  7 +++++--
 drivers/net/mlx5/mlx5_rxtx.c | 44 +++++++++++++++++++++++++-------------------
 drivers/net/mlx5/mlx5_rxtx.h |  5 ++---
 drivers/net/mlx5/mlx5_txq.c  | 21 ++++++++++++---------
 4 files changed, 44 insertions(+), 33 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h
index 8d2ec7a..cc2a6f3 100644
--- a/drivers/net/mlx5/mlx5_defs.h
+++ b/drivers/net/mlx5/mlx5_defs.h
@@ -48,8 +48,11 @@
 /* Maximum number of special flows. */
 #define MLX5_MAX_SPECIAL_FLOWS 4
 
-/* Request send completion once in every 64 sends, might be less. */
-#define MLX5_PMD_TX_PER_COMP_REQ 64
+/*
+ * Request TX completion every time descriptors reach this threshold since
+ * the previous request. Must be a power of two for performance reasons.
+ */
+#define MLX5_TX_COMP_THRESH 32
 
 /* RSS Indirection table size. */
 #define RSS_INDIRECTION_TABLE_SIZE 256
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 43236f5..9d992c3 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -156,9 +156,6 @@ check_cqe64(volatile struct mlx5_cqe64 *cqe,
  * Manage TX completions.
  *
  * When sending a burst, mlx5_tx_burst() posts several WRs.
- * To improve performance, a completion event is only required once every
- * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
- * for other WRs, but this information would not be used anyway.
  *
  * @param txq
  *   Pointer to TX queue structure.
@@ -172,14 +169,16 @@ txq_complete(struct txq *txq)
 	uint16_t elts_free = txq->elts_tail;
 	uint16_t elts_tail;
 	uint16_t cq_ci = txq->cq_ci;
-	unsigned int wqe_ci = (unsigned int)-1;
+	volatile struct mlx5_cqe64 *cqe = NULL;
+	volatile union mlx5_wqe *wqe;
 
 	do {
-		unsigned int idx = cq_ci & cqe_cnt;
-		volatile struct mlx5_cqe64 *cqe = &(*txq->cqes)[idx].cqe64;
+		volatile struct mlx5_cqe64 *tmp;
 
-		if (check_cqe64(cqe, cqe_n, cq_ci) == 1)
+		tmp = &(*txq->cqes)[cq_ci & cqe_cnt].cqe64;
+		if (check_cqe64(tmp, cqe_n, cq_ci))
 			break;
+		cqe = tmp;
 #ifndef NDEBUG
 		if (MLX5_CQE_FORMAT(cqe->op_own) == MLX5_COMPRESSED) {
 			if (!check_cqe64_seen(cqe))
@@ -193,14 +192,15 @@ txq_complete(struct txq *txq)
 			return;
 		}
 #endif /* NDEBUG */
-		wqe_ci = ntohs(cqe->wqe_counter);
 		++cq_ci;
 	} while (1);
-	if (unlikely(wqe_ci == (unsigned int)-1))
+	if (unlikely(cqe == NULL))
 		return;
+	wqe = &(*txq->wqes)[htons(cqe->wqe_counter) & (txq->wqe_n - 1)];
+	elts_tail = wqe->wqe.ctrl.data[3];
+	assert(elts_tail < txq->wqe_n);
 	/* Free buffers. */
-	elts_tail = (wqe_ci + 1) & (elts_n - 1);
-	do {
+	while (elts_free != elts_tail) {
 		struct rte_mbuf *elt = (*txq->elts)[elts_free];
 		unsigned int elts_free_next =
 			(elts_free + 1) & (elts_n - 1);
@@ -216,7 +216,7 @@ txq_complete(struct txq *txq)
 		/* Only one segment needs to be freed. */
 		rte_pktmbuf_free_seg(elt);
 		elts_free = elts_free_next;
-	} while (elts_free != elts_tail);
+	}
 	txq->cq_ci = cq_ci;
 	txq->elts_tail = elts_tail;
 	/* Update the consumer index. */
@@ -437,6 +437,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	const unsigned int elts_n = txq->elts_n;
 	unsigned int i;
 	unsigned int max;
+	unsigned int comp;
 	volatile union mlx5_wqe *wqe;
 	struct rte_mbuf *buf;
 
@@ -486,13 +487,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 					    buf->vlan_tci);
 		else
 			mlx5_wqe_write(txq, wqe, addr, length, lkey);
-		/* Request completion if needed. */
-		if (unlikely(--txq->elts_comp == 0)) {
-			wqe->wqe.ctrl.data[2] = htonl(8);
-			txq->elts_comp = txq->elts_comp_cd_init;
-		} else {
-			wqe->wqe.ctrl.data[2] = 0;
-		}
+		wqe->wqe.ctrl.data[2] = 0;
 		/* Should we enable HW CKSUM offload */
 		if (buf->ol_flags &
 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
@@ -512,6 +507,17 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	/* Take a shortcut if nothing must be sent. */
 	if (unlikely(i == 0))
 		return 0;
+	/* Check whether completion threshold has been reached. */
+	comp = txq->elts_comp + i;
+	if (comp >= MLX5_TX_COMP_THRESH) {
+		/* Request completion on last WQE. */
+		wqe->wqe.ctrl.data[2] = htonl(8);
+		/* Save elts_head in unused "immediate" field of WQE. */
+		wqe->wqe.ctrl.data[3] = elts_head;
+		txq->elts_comp = 0;
+	} else {
+		txq->elts_comp = comp;
+	}
 #ifdef MLX5_PMD_SOFT_COUNTERS
 	/* Increment sent packets counter. */
 	txq->stats.opackets += i;
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index 77b0fde..f900e65 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -238,8 +238,7 @@ struct hash_rxq {
 struct txq {
 	uint16_t elts_head; /* Current index in (*elts)[]. */
 	uint16_t elts_tail; /* First element awaiting completion. */
-	uint16_t elts_comp_cd_init; /* Initial value for countdown. */
-	uint16_t elts_comp; /* Elements before asking a completion. */
+	uint16_t elts_comp; /* Counter since last completion request. */
 	uint16_t elts_n; /* (*elts)[] length. */
 	uint16_t cq_ci; /* Consumer index for completion queue. */
 	uint16_t cqe_n; /* Number of CQ elements. */
@@ -247,6 +246,7 @@ struct txq {
 	uint16_t wqe_n; /* Number of WQ elements. */
 	uint16_t bf_offset; /* Blueflame offset. */
 	uint16_t bf_buf_size; /* Blueflame size. */
+	uint32_t qp_num_8s; /* QP number shifted by 8. */
 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
 	volatile union mlx5_wqe (*wqes)[]; /* Work queue. */
 	volatile uint32_t *qp_db; /* Work queue doorbell. */
@@ -259,7 +259,6 @@ struct txq {
 	} mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
 	struct rte_mbuf *(*elts)[]; /* TX elements. */
 	struct mlx5_txq_stats stats; /* TX queue counters. */
-	uint32_t qp_num_8s; /* QP number shifted by 8. */
 } __rte_cache_aligned;
 
 /* TX queue control descriptor. */
diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
index 22e9bae..7b2dc7c 100644
--- a/drivers/net/mlx5/mlx5_txq.c
+++ b/drivers/net/mlx5/mlx5_txq.c
@@ -89,6 +89,7 @@ txq_alloc_elts(struct txq_ctrl *txq_ctrl, unsigned int elts_n)
 	DEBUG("%p: allocated and configured %u WRs", (void *)txq_ctrl, elts_n);
 	txq_ctrl->txq.elts_head = 0;
 	txq_ctrl->txq.elts_tail = 0;
+	txq_ctrl->txq.elts_comp = 0;
 }
 
 /**
@@ -108,6 +109,7 @@ txq_free_elts(struct txq_ctrl *txq_ctrl)
 	DEBUG("%p: freeing WRs", (void *)txq_ctrl);
 	txq_ctrl->txq.elts_head = 0;
 	txq_ctrl->txq.elts_tail = 0;
+	txq_ctrl->txq.elts_comp = 0;
 
 	while (elts_tail != elts_head) {
 		struct rte_mbuf *elt = (*elts)[elts_tail];
@@ -274,15 +276,8 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
 		goto error;
 	}
 	(void)conf; /* Thresholds configuration (ignored). */
+	assert(desc > MLX5_TX_COMP_THRESH);
 	tmpl.txq.elts_n = desc;
-	/*
-	 * Request send completion every MLX5_PMD_TX_PER_COMP_REQ packets or
-	 * at least 4 times per ring.
-	 */
-	tmpl.txq.elts_comp_cd_init =
-		((MLX5_PMD_TX_PER_COMP_REQ < (desc / 4)) ?
-		 MLX5_PMD_TX_PER_COMP_REQ : (desc / 4));
-	tmpl.txq.elts_comp = tmpl.txq.elts_comp_cd_init;
 	/* MRs will be registered in mp2mr[] later. */
 	attr.rd = (struct ibv_exp_res_domain_init_attr){
 		.comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
@@ -302,7 +297,8 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
 		.res_domain = tmpl.rd,
 	};
 	tmpl.cq = ibv_exp_create_cq(priv->ctx,
-				    (desc / tmpl.txq.elts_comp_cd_init) - 1,
+				    (((desc / MLX5_TX_COMP_THRESH) - 1) ?
+				     ((desc / MLX5_TX_COMP_THRESH) - 1) : 1),
 				    NULL, NULL, 0, &attr.cq);
 	if (tmpl.cq == NULL) {
 		ret = ENOMEM;
@@ -454,6 +450,13 @@ mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		return -E_RTE_SECONDARY;
 
 	priv_lock(priv);
+	if (desc <= MLX5_TX_COMP_THRESH) {
+		WARN("%p: number of descriptors requested for TX queue %u"
+		     " must be higher than MLX5_TX_COMP_THRESH, using"
+		     " %u instead of %u",
+		     (void *)dev, idx, MLX5_TX_COMP_THRESH + 1, desc);
+		desc = MLX5_TX_COMP_THRESH + 1;
+	}
 	if (!rte_is_power_of_2(desc)) {
 		desc = 1 << log2above(desc);
 		WARN("%p: increased number of descriptors in TX queue %u"
-- 
2.1.4