commit 6b9fde2b470326a240154f2c8b6ad16cb86c8371
Author: Ryan Stone <rstone@FreeBSD.org>
Date:   Thu Aug 4 14:11:36 2016

    Move debug counters to per-ring structure
    
    Currently the mlx4_en driver increments a number of port-wide
    counters in the rx and tx paths.  This causes the cache line
    containing those counters to ping-pong between CPUs when more
    than one ring is being used simultaneously, lowering perf.
    
    The driver already has the infrastructure to maintain separate
    per-ring counters and then aggregate them together periodically,
    so make use of that to fix the cache problems.

diff --git a/sys/ofed/drivers/net/mlx4/en_port.c b/sys/ofed/drivers/net/mlx4/en_port.c
index 2a1c4ba..ec7d3f9 100644
--- a/sys/ofed/drivers/net/mlx4/en_port.c
+++ b/sys/ofed/drivers/net/mlx4/en_port.c
@@ -191,11 +191,14 @@ int mlx4_en_DUMP_ETH_STATS(struct mlx4_en_dev *mdev, u8 port, u8 reset)
 	priv->port_stats.tx_chksum_offload = 0;
 	priv->port_stats.queue_stopped = 0;
 	priv->port_stats.wake_queue = 0;
+	priv->port_stats.oversized_packets = 0;
+	priv->port_stats.tso_packets = 0;
 	for (i = 0; i < priv->tx_ring_num; i++) {
 		priv->port_stats.tx_chksum_offload += priv->tx_ring[i]->tx_csum;
 		priv->port_stats.queue_stopped += priv->tx_ring[i]->queue_stopped;
 		priv->port_stats.wake_queue += priv->tx_ring[i]->wake_queue;
 		priv->port_stats.oversized_packets += priv->tx_ring[i]->oversized_packets;
+		priv->port_stats.tso_packets += priv->tx_ring[i]->tso_packets;
 	}
 	/* RX Statistics */
 	priv->pkstats.rx_packets = be64_to_cpu(mlx4_en_stats->RTOT_prio_0) +
diff --git a/sys/ofed/drivers/net/mlx4/en_rx.c b/sys/ofed/drivers/net/mlx4/en_rx.c
index 2711b22..954278c 100644
--- a/sys/ofed/drivers/net/mlx4/en_rx.c
+++ b/sys/ofed/drivers/net/mlx4/en_rx.c
@@ -630,7 +630,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 		    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
 		    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
 		    (cqe->checksum == cpu_to_be16(0xffff))) {
-			priv->port_stats.rx_chksum_good++;
+			ring->csum_ok++;
 			mb->m_pkthdr.csum_flags =
 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
@@ -654,7 +654,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 			INC_PERF_COUNTER(priv->pstats.lro_misses);
 		} else {
 			mb->m_pkthdr.csum_flags = 0;
-			priv->port_stats.rx_chksum_none++;
+			ring->csum_none++;
 		}
 
 		/* Push it up the stack */
diff --git a/sys/ofed/drivers/net/mlx4/en_tx.c b/sys/ofed/drivers/net/mlx4/en_tx.c
index 9090c51..ae21d33 100644
--- a/sys/ofed/drivers/net/mlx4/en_tx.c
+++ b/sys/ofed/drivers/net/mlx4/en_tx.c
@@ -448,7 +448,6 @@ static int mlx4_en_process_tx_cq(struct net_device *dev,
 		if (atomic_fetchadd_int(&priv->blocked, -1) == 1)
 			atomic_clear_int(&dev->if_drv_flags ,IFF_DRV_OACTIVE);
 		ring->wake_queue++;
-		priv->port_stats.wake_queue++;
 	}
 	return (0);
 }
@@ -714,7 +713,6 @@ static int mlx4_en_xmit(struct mlx4_en_priv *priv, int tx_ind, struct mbuf **mbp
 			atomic_set_int(&ifp->if_drv_flags, IFF_DRV_OACTIVE);
 			priv->port_stats.queue_stopped++;
 		ring->blocked = 1;
-		priv->port_stats.queue_stopped++;
 		ring->queue_stopped++;
 
 		/* Use interrupts to find out when queue opened */
@@ -758,7 +756,6 @@ static int mlx4_en_xmit(struct mlx4_en_priv *priv, int tx_ind, struct mbuf **mbp
 
 	/* do statistics */
 	if (likely(tx_desc->ctrl.srcrb_flags != CTRL_FLAGS)) {
-		priv->port_stats.tx_chksum_offload++;
 		ring->tx_csum++;
 	}
 
@@ -796,7 +793,7 @@ static int mlx4_en_xmit(struct mlx4_en_priv *priv, int tx_ind, struct mbuf **mbp
 			num_pkts = DIV_ROUND_UP(payload_len, mss);
 		ring->bytes += payload_len + (num_pkts * ihs);
 		ring->packets += num_pkts;
-		priv->port_stats.tso_packets++;
+		ring->tso_packets++;
 		/* store pointer to inline header */
 		dseg_inline = dseg;
 		/* copy data inline */
diff --git a/sys/ofed/drivers/net/mlx4/mlx4_en.h b/sys/ofed/drivers/net/mlx4/mlx4_en.h
index 17afa9e..f6ea011 100644
--- a/sys/ofed/drivers/net/mlx4/mlx4_en.h
+++ b/sys/ofed/drivers/net/mlx4/mlx4_en.h
@@ -278,6 +278,7 @@ struct mlx4_en_tx_ring {
 	unsigned long queue_stopped;
 	unsigned long oversized_packets;
 	unsigned long wake_queue;
+	unsigned long tso_packets;
 	struct mlx4_bf bf;
 	bool bf_enabled;
 	int hwtstamp_tx_type;