[dpdk-dev] [PATCH 3/3] net/mlx5: fix the Tx completion request generation

[Viacheslav Ovsiienko <viacheslavo@mellanox.com>]

The packets transmitting in mlx5 is performed by building
Tx descriptors (WQEs) and sending last ones to the NIC.
The descriptor can contain the special flags, telling the NIC
to generate Tx completion notification (CQEs). At the beginning
of tx_burst() routine PMD checks whether there are some Tx
completions and frees the transmitted packet buffers.

The flags to request completion generation must be set once
per specified amount of packets to provide uniform stream
of completions and freeing the Tx queue in unifirm fashion.
The previous implementation sets the completion request
generation once per burst, if burst size if big enough it may
latency in CQE generation and freeing large amount of buffers
in tx_burst routine on multiple completions which also
affects the latency and even causes the Tx queue overflow
and Tx drops.

This patches enforces the completion request will be set
in the exact Tx descriptor if specified amount of packets
is already sent.

Fixes: 18a1c20044c0 ("net/mlx5: implement Tx burst template")

Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
---
 drivers/net/mlx5/mlx5_defs.h |  2 +-
 drivers/net/mlx5/mlx5_prm.h  | 17 +++++++-----
 drivers/net/mlx5/mlx5_rxtx.c | 64 +++++++++++++++++++++++++++++---------------
 3 files changed, 55 insertions(+), 28 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h
index 461e916..d7440fd 100644
--- a/drivers/net/mlx5/mlx5_defs.h
+++ b/drivers/net/mlx5/mlx5_defs.h
@@ -28,7 +28,7 @@
  * 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
+#define MLX5_TX_COMP_THRESH 32u
 
 /*
  * Request TX completion every time the total number of WQEBBs used for inlining
diff --git a/drivers/net/mlx5/mlx5_prm.h b/drivers/net/mlx5/mlx5_prm.h
index 32bc7a6..89548d4 100644
--- a/drivers/net/mlx5/mlx5_prm.h
+++ b/drivers/net/mlx5/mlx5_prm.h
@@ -72,7 +72,7 @@
  * boundary with accounting the title Control and Ethernet
  * segments.
  */
-#define MLX5_EMPW_DEF_INLINE_LEN (3U * MLX5_WQE_SIZE + \
+#define MLX5_EMPW_DEF_INLINE_LEN (3u * MLX5_WQE_SIZE + \
 				  MLX5_DSEG_MIN_INLINE_SIZE - \
 				  MLX5_WQE_DSEG_SIZE)
 /*
@@ -90,11 +90,16 @@
  * If there are no enough resources to built minimal
  * EMPW the sending loop exits.
  */
-#define MLX5_EMPW_MIN_PACKETS (2 + 3 * 4)
-#define MLX5_EMPW_MAX_PACKETS ((MLX5_WQE_SIZE_MAX - \
-				MLX5_WQE_CSEG_SIZE - \
-				MLX5_WQE_ESEG_SIZE) / \
-				MLX5_WSEG_SIZE)
+#define MLX5_EMPW_MIN_PACKETS (2u + 3u * 4u)
+/*
+ * Maximal amount of packets to be sent with EMPW.
+ * This value is not recommended to exceed MLX5_TX_COMP_THRESH,
+ * otherwise there might be up to MLX5_EMPW_MAX_PACKETS mbufs
+ * without CQE generation request, being multiplied by
+ * MLX5_TX_COMP_MAX_CQE it may cause significant latency
+ * in tx burst routine at the moment of freeing multiple mbufs.
+ */
+#define MLX5_EMPW_MAX_PACKETS MLX5_TX_COMP_THRESH
 /*
  * Default packet length threshold to be inlined with
  * ordinary SEND. Inlining saves the MR key search
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index c2b93c6..5984c50 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -2063,8 +2063,6 @@ enum mlx5_txcmp_code {
  *
  * @param txq
  *   Pointer to TX queue structure.
- * @param n_mbuf
- *   Number of mbuf not stored yet in elts array.
  * @param loc
  *   Pointer to burst routine local context.
  * @param olx
@@ -2073,18 +2071,23 @@ enum mlx5_txcmp_code {
  */
 static __rte_always_inline void
 mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq,
-			   unsigned int n_mbuf,
 			   struct mlx5_txq_local *restrict loc,
-			   unsigned int olx __rte_unused)
+			   unsigned int olx)
 {
-	uint16_t head = txq->elts_head + n_mbuf;
+	uint16_t head = txq->elts_head;
+	unsigned int part;
 
+	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc->pkts_sent -
+		(MLX5_TXOFF_CONFIG(MULTI) ? loc->pkts_copy : 0);
+	head += part;
 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
-	    (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres) {
+	     (MLX5_TXOFF_CONFIG(INLINE) &&
+	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
 		volatile struct mlx5_wqe *last = loc->wqe_last;
 
 		txq->elts_comp = head;
-		txq->wqe_comp = txq->wqe_ci;
+		if (MLX5_TXOFF_CONFIG(INLINE))
+			txq->wqe_comp = txq->wqe_ci;
 		/* Request unconditional completion on last WQE. */
 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
 					    MLX5_COMP_MODE_OFFSET);
@@ -3023,6 +3026,8 @@ enum mlx5_txcmp_code {
 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
 	txq->wqe_ci += (ds + 3) / 4;
 	loc->wqe_free -= (ds + 3) / 4;
+	/* Request CQE generation if limits are reached. */
+	mlx5_tx_request_completion(txq, loc, olx);
 	return MLX5_TXCMP_CODE_MULTI;
 }
 
@@ -3131,6 +3136,8 @@ enum mlx5_txcmp_code {
 	} while (true);
 	txq->wqe_ci += (ds + 3) / 4;
 	loc->wqe_free -= (ds + 3) / 4;
+	/* Request CQE generation if limits are reached. */
+	mlx5_tx_request_completion(txq, loc, olx);
 	return MLX5_TXCMP_CODE_MULTI;
 }
 
@@ -3287,6 +3294,8 @@ enum mlx5_txcmp_code {
 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
 	txq->wqe_ci += (ds + 3) / 4;
 	loc->wqe_free -= (ds + 3) / 4;
+	/* Request CQE generation if limits are reached. */
+	mlx5_tx_request_completion(txq, loc, olx);
 	return MLX5_TXCMP_CODE_MULTI;
 }
 
@@ -3496,6 +3505,8 @@ enum mlx5_txcmp_code {
 		--loc->elts_free;
 		++loc->pkts_sent;
 		--pkts_n;
+		/* Request CQE generation if limits are reached. */
+		mlx5_tx_request_completion(txq, loc, olx);
 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
 			return MLX5_TXCMP_CODE_EXIT;
 		loc->mbuf = *pkts++;
@@ -3637,7 +3648,7 @@ enum mlx5_txcmp_code {
 		   struct mlx5_txq_local *restrict loc,
 		   unsigned int ds,
 		   unsigned int slen,
-		   unsigned int olx __rte_unused)
+		   unsigned int olx)
 {
 	assert(!MLX5_TXOFF_CONFIG(INLINE));
 #ifdef MLX5_PMD_SOFT_COUNTERS
@@ -3652,6 +3663,8 @@ enum mlx5_txcmp_code {
 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
 	txq->wqe_ci += (ds + 3) / 4;
 	loc->wqe_free -= (ds + 3) / 4;
+	/* Request CQE generation if limits are reached. */
+	mlx5_tx_request_completion(txq, loc, olx);
 }
 
 /*
@@ -3694,6 +3707,8 @@ enum mlx5_txcmp_code {
 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
 	txq->wqe_ci += (len + 3) / 4;
 	loc->wqe_free -= (len + 3) / 4;
+	/* Request CQE generation if limits are reached. */
+	mlx5_tx_request_completion(txq, loc, olx);
 }
 
 /**
@@ -3865,6 +3880,7 @@ enum mlx5_txcmp_code {
 				if (unlikely(!loc->elts_free ||
 					     !loc->wqe_free))
 					return MLX5_TXCMP_CODE_EXIT;
+				pkts_n -= part;
 				goto next_empw;
 			}
 			/* Packet attributes match, continue the same eMPW. */
@@ -3884,6 +3900,8 @@ enum mlx5_txcmp_code {
 		txq->wqe_ci += (2 + part + 3) / 4;
 		loc->wqe_free -= (2 + part + 3) / 4;
 		pkts_n -= part;
+		/* Request CQE generation if limits are reached. */
+		mlx5_tx_request_completion(txq, loc, olx);
 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
 			return MLX5_TXCMP_CODE_EXIT;
 		loc->mbuf = *pkts++;
@@ -3922,10 +3940,14 @@ enum mlx5_txcmp_code {
 		struct mlx5_wqe_dseg *restrict dseg;
 		struct mlx5_wqe_eseg *restrict eseg;
 		enum mlx5_txcmp_code ret;
-		unsigned int room, part;
+		unsigned int room, part, nlim;
 		unsigned int slen = 0;
 
-next_empw:
+		/*
+		 * Limits the amount of packets in one WQE
+		 * to improve CQE latency generation.
+		 */
+		nlim = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS);
 		/* Check whether we have minimal amount WQEs */
 		if (unlikely(loc->wqe_free <
 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
@@ -4044,12 +4066,6 @@ enum mlx5_txcmp_code {
 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
 				return MLX5_TXCMP_CODE_EXIT;
 			}
-			/* Check if we have minimal room left. */
-			if (room < MLX5_WQE_DSEG_SIZE) {
-				part -= room;
-				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
-				goto next_empw;
-			}
 			loc->mbuf = *pkts++;
 			if (likely(pkts_n > 1))
 				rte_prefetch0(*pkts);
@@ -4089,6 +4105,10 @@ enum mlx5_txcmp_code {
 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
 				return MLX5_TXCMP_CODE_ERROR;
 			}
+			/* Check if we have minimal room left. */
+			nlim--;
+			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
+				break;
 			/*
 			 * Check whether packet parameters coincide
 			 * within assumed eMPW batch:
@@ -4114,7 +4134,7 @@ enum mlx5_txcmp_code {
 		if (unlikely(!loc->elts_free ||
 			     !loc->wqe_free))
 			return MLX5_TXCMP_CODE_EXIT;
-		goto next_empw;
+		/* Continue the loop with new eMPW session. */
 	}
 	assert(false);
 }
@@ -4355,6 +4375,8 @@ enum mlx5_txcmp_code {
 		}
 		++loc->pkts_sent;
 		--pkts_n;
+		/* Request CQE generation if limits are reached. */
+		mlx5_tx_request_completion(txq, loc, olx);
 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
 			return MLX5_TXCMP_CODE_EXIT;
 		loc->mbuf = *pkts++;
@@ -4630,9 +4652,6 @@ enum mlx5_txcmp_code {
 	/* Take a shortcut if nothing is sent. */
 	if (unlikely(loc.pkts_sent == 0))
 		return 0;
-	/* Not all of the mbufs may be stored into elts yet. */
-	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
-	mlx5_tx_request_completion(txq, part, &loc, olx);
 	/*
 	 * Ring QP doorbell immediately after WQE building completion
 	 * to improve latencies. The pure software related data treatment
@@ -4640,10 +4659,13 @@ enum mlx5_txcmp_code {
 	 * processed in this thread only by the polling.
 	 */
 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, 0);
+	/* Not all of the mbufs may be stored into elts yet. */
+	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent -
+		(MLX5_TXOFF_CONFIG(MULTI) ? loc.pkts_copy : 0);
 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
 		/*
 		 * There are some single-segment mbufs not stored in elts.
-		 * It can be only if last packet was single-segment.
+		 * It can be only if the last packet was single-segment.
 		 * The copying is gathered into one place due to it is
 		 * a good opportunity to optimize that with SIMD.
 		 * Unfortunately if inlining is enabled the gaps in
-- 
1.8.3.1