[PATCH net-next 00/10] mlx4 XDP performance improvements

[PATCH net-next 00/10] mlx4 XDP performance improvements

[Tariq Toukan]

Hi Dave,

This patchset contains data-path improvements, mainly for XDP_DROP
and XDP_TX cases.

Main patches:
* Patch 2 by Saeed allows enabling optimized A0 RX steering (in HW) when
  setting a single RX ring.
  With this configuration, HW packet-rate dramatically improves,
  reaching 28.1 Mpps in XDP_DROP case for both IPv4 (37% gain)
  and IPv6 (53% gain).
* Patch 6 enhances the XDP xmit function. Among other changes, now we
  ring one doorbell per NAPI. Patch gives 17% gain in XDP_TX case.
* Patch 7 obsoletes the NAPI of XDP_TX completion queue and integrates its
  poll into the respective RX NAPI. Patch gives 15% gain in XDP_TX case.

Series generated against net-next commit:
f7aec129a356 rxrpc: Cache the congestion window setting

Thanks,
Tariq.


Saeed Mahameed (1):
  net/mlx4_en: Optimized single ring steering

Tariq Toukan (9):
  net/mlx4_en: Remove unused argument in TX datapath function
  net/mlx4_en: Improve receive data-path
  net/mlx4_en: Improve transmit CQ polling
  net/mlx4_en: Improve stack xmit function
  net/mlx4_en: Improve XDP xmit function
  net/mlx4_en: Poll XDP TX completion queue in RX NAPI
  net/mlx4_en: Increase default TX ring size
  net/mlx4_en: Replace TXBB_SIZE multiplications with shift operations
  net/mlx4_en: Refactor mlx4_en_free_tx_desc

 drivers/net/ethernet/mellanox/mlx4/en_cq.c     |  25 +-
 drivers/net/ethernet/mellanox/mlx4/en_main.c   |   6 +-
 drivers/net/ethernet/mellanox/mlx4/en_netdev.c |  20 +-
 drivers/net/ethernet/mellanox/mlx4/en_rx.c     | 145 ++++++++----
 drivers/net/ethernet/mellanox/mlx4/en_tx.c     | 305 ++++++++++++-------------
 drivers/net/ethernet/mellanox/mlx4/main.c      |   4 +-
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h   |  23 +-
 7 files changed, 293 insertions(+), 235 deletions(-)

-- 
1.8.3.1

[PATCH net-next 01/10] net/mlx4_en: Remove unused argument in TX datapath function

[Tariq Toukan]

Remove owner argument, as it is obsolete and unused.
This also saves the overhead of calculating its value in data-path.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 10 ++++------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  6 +++---
 2 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 6ffd1849a604..37386abea54c 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -265,7 +265,7 @@ static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
 
 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 			 struct mlx4_en_tx_ring *ring,
-			 int index, u8 owner, u64 timestamp,
+			 int index, u64 timestamp,
 			 int napi_mode)
 {
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
@@ -344,7 +344,7 @@ u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 
 u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
 			    struct mlx4_en_tx_ring *ring,
-			    int index, u8 owner, u64 timestamp,
+			    int index, u64 timestamp,
 			    int napi_mode)
 {
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
@@ -381,8 +381,7 @@ int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
 	while (ring->cons != ring->prod) {
 		ring->last_nr_txbb = ring->free_tx_desc(priv, ring,
 						ring->cons & ring->size_mask,
-						!!(ring->cons & ring->size), 0,
-						0 /* Non-NAPI caller */);
+						0, 0 /* Non-NAPI caller */);
 		ring->cons += ring->last_nr_txbb;
 		cnt++;
 	}
@@ -464,8 +463,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 			/* free next descriptor */
 			last_nr_txbb = ring->free_tx_desc(
 					priv, ring, ring_index,
-					!!((ring_cons + txbbs_skipped) &
-					ring->size), timestamp, napi_budget);
+					timestamp, napi_budget);
 
 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
 					  !!((ring_cons + txbbs_stamp) &
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 8c4f63946b14..08b2906a98af 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -276,7 +276,7 @@ struct mlx4_en_tx_ring {
 	struct netdev_queue	*tx_queue;
 	u32			(*free_tx_desc)(struct mlx4_en_priv *priv,
 						struct mlx4_en_tx_ring *ring,
-						int index, u8 owner,
+						int index,
 						u64 timestamp, int napi_mode);
 	struct mlx4_en_rx_ring	*recycle_ring;
 
@@ -723,11 +723,11 @@ int mlx4_en_process_rx_cq(struct net_device *dev,
 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget);
 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 			 struct mlx4_en_tx_ring *ring,
-			 int index, u8 owner, u64 timestamp,
+			 int index, u64 timestamp,
 			 int napi_mode);
 u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
 			    struct mlx4_en_tx_ring *ring,
-			    int index, u8 owner, u64 timestamp,
+			    int index, u64 timestamp,
 			    int napi_mode);
 void mlx4_en_fill_qp_context(struct mlx4_en_priv *priv, int size, int stride,
 		int is_tx, int rss, int qpn, int cqn, int user_prio,
-- 
1.8.3.1

[PATCH net-next 02/10] net/mlx4_en: Optimized single ring steering

[Tariq Toukan]

From: Saeed Mahameed <saeedm@mellanox.com>

Avoid touching RX QP RSS context when loading with only
one RX ring, to allow optimized A0 RX steering.

Enable by:
- loading mlx4_core with module param: log_num_mgm_entry_size = -6.
- then: ethtool -L <interface> rx 1

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

XDP_DROP packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 20.5 Mpps | 28.1 Mpps |  37% |
IPv6 | 18.4 Mpps | 28.1 Mpps |  53% |
-------------------------------------

Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_main.c   |  6 ++--
 drivers/net/ethernet/mellanox/mlx4/en_netdev.c | 12 ++++---
 drivers/net/ethernet/mellanox/mlx4/en_rx.c     | 48 ++++++++++++++++++++------
 drivers/net/ethernet/mellanox/mlx4/main.c      |  4 +--
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h   |  2 +-
 5 files changed, 50 insertions(+), 22 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_main.c b/drivers/net/ethernet/mellanox/mlx4/en_main.c
index d94f981eafc4..56cdf38d150e 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_main.c
@@ -125,9 +125,9 @@ void mlx4_en_update_loopback_state(struct net_device *dev,
 		priv->flags |= MLX4_EN_FLAG_ENABLE_HW_LOOPBACK;
 
 	mutex_lock(&priv->mdev->state_lock);
-	if (priv->mdev->dev->caps.flags2 &
-	    MLX4_DEV_CAP_FLAG2_UPDATE_QP_SRC_CHECK_LB &&
-	    priv->rss_map.indir_qp.qpn) {
+	if ((priv->mdev->dev->caps.flags2 &
+	     MLX4_DEV_CAP_FLAG2_UPDATE_QP_SRC_CHECK_LB) &&
+	    priv->rss_map.indir_qp && priv->rss_map.indir_qp->qpn) {
 		int i;
 		int err = 0;
 		int loopback = !!(features & NETIF_F_LOOPBACK);
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
index c1de75fc399a..51ce111b719e 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
@@ -596,6 +596,8 @@ static int mlx4_en_get_qp(struct mlx4_en_priv *priv)
 		return err;
 	}
 
+	en_info(priv, "Steering Mode %d\n", dev->caps.steering_mode);
+
 	if (dev->caps.steering_mode == MLX4_STEERING_MODE_A0) {
 		int base_qpn = mlx4_get_base_qpn(dev, priv->port);
 		*qpn = base_qpn + index;
@@ -1010,7 +1012,7 @@ static void mlx4_en_do_multicast(struct mlx4_en_priv *priv,
 				memcpy(&mc_list[10], mclist->addr, ETH_ALEN);
 				mc_list[5] = priv->port;
 				err = mlx4_multicast_detach(mdev->dev,
-							    &priv->rss_map.indir_qp,
+							    priv->rss_map.indir_qp,
 							    mc_list,
 							    MLX4_PROT_ETH,
 							    mclist->reg_id);
@@ -1032,7 +1034,7 @@ static void mlx4_en_do_multicast(struct mlx4_en_priv *priv,
 				/* needed for B0 steering support */
 				mc_list[5] = priv->port;
 				err = mlx4_multicast_attach(mdev->dev,
-							    &priv->rss_map.indir_qp,
+							    priv->rss_map.indir_qp,
 							    mc_list,
 							    priv->port, 0,
 							    MLX4_PROT_ETH,
@@ -1742,7 +1744,7 @@ int mlx4_en_start_port(struct net_device *dev)
 	/* Attach rx QP to bradcast address */
 	eth_broadcast_addr(&mc_list[10]);
 	mc_list[5] = priv->port; /* needed for B0 steering support */
-	if (mlx4_multicast_attach(mdev->dev, &priv->rss_map.indir_qp, mc_list,
+	if (mlx4_multicast_attach(mdev->dev, priv->rss_map.indir_qp, mc_list,
 				  priv->port, 0, MLX4_PROT_ETH,
 				  &priv->broadcast_id))
 		mlx4_warn(mdev, "Failed Attaching Broadcast\n");
@@ -1866,12 +1868,12 @@ void mlx4_en_stop_port(struct net_device *dev, int detach)
 	/* Detach All multicasts */
 	eth_broadcast_addr(&mc_list[10]);
 	mc_list[5] = priv->port; /* needed for B0 steering support */
-	mlx4_multicast_detach(mdev->dev, &priv->rss_map.indir_qp, mc_list,
+	mlx4_multicast_detach(mdev->dev, priv->rss_map.indir_qp, mc_list,
 			      MLX4_PROT_ETH, priv->broadcast_id);
 	list_for_each_entry(mclist, &priv->curr_list, list) {
 		memcpy(&mc_list[10], mclist->addr, ETH_ALEN);
 		mc_list[5] = priv->port;
-		mlx4_multicast_detach(mdev->dev, &priv->rss_map.indir_qp,
+		mlx4_multicast_detach(mdev->dev, priv->rss_map.indir_qp,
 				      mc_list, MLX4_PROT_ETH, mclist->reg_id);
 		if (mclist->tunnel_reg_id)
 			mlx4_flow_detach(mdev->dev, mclist->tunnel_reg_id);
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index 77abd1813047..c4edae854f1b 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
@@ -1099,11 +1099,14 @@ int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
 	int i, qpn;
 	int err = 0;
 	int good_qps = 0;
+	u8 flags;
 
 	en_dbg(DRV, priv, "Configuring rss steering\n");
+
+	flags = priv->rx_ring_num == 1 ? MLX4_RESERVE_A0_QP : 0;
 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
 				    priv->rx_ring_num,
-				    &rss_map->base_qpn, 0);
+				    &rss_map->base_qpn, flags);
 	if (err) {
 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
 		return err;
@@ -1120,13 +1123,28 @@ int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
 		++good_qps;
 	}
 
+	if (priv->rx_ring_num == 1) {
+		rss_map->indir_qp = &rss_map->qps[0];
+		priv->base_qpn = rss_map->indir_qp->qpn;
+		en_info(priv, "Optimized Non-RSS steering\n");
+		return 0;
+	}
+
+	rss_map->indir_qp = kzalloc(sizeof(*rss_map->indir_qp), GFP_KERNEL);
+	if (!rss_map->indir_qp) {
+		err = -ENOMEM;
+		goto rss_err;
+	}
+
 	/* Configure RSS indirection qp */
-	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp, GFP_KERNEL);
+	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, rss_map->indir_qp,
+			    GFP_KERNEL);
 	if (err) {
 		en_err(priv, "Failed to allocate RSS indirection QP\n");
 		goto rss_err;
 	}
-	rss_map->indir_qp.event = mlx4_en_sqp_event;
+
+	rss_map->indir_qp->event = mlx4_en_sqp_event;
 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
 				priv->rx_ring[0]->cqn, -1, &context);
 
@@ -1164,8 +1182,9 @@ int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
 		err = -EINVAL;
 		goto indir_err;
 	}
+
 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
-			       &rss_map->indir_qp, &rss_map->indir_state);
+			       rss_map->indir_qp, &rss_map->indir_state);
 	if (err)
 		goto indir_err;
 
@@ -1173,9 +1192,11 @@ int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
 
 indir_err:
 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
-		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
-	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
-	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
+		       MLX4_QP_STATE_RST, NULL, 0, 0, rss_map->indir_qp);
+	mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
+	mlx4_qp_free(mdev->dev, rss_map->indir_qp);
+	kfree(rss_map->indir_qp);
+	rss_map->indir_qp = NULL;
 rss_err:
 	for (i = 0; i < good_qps; i++) {
 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
@@ -1193,10 +1214,15 @@ void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
 	int i;
 
-	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
-		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
-	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
-	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
+	if (priv->rx_ring_num > 1) {
+		mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
+			       MLX4_QP_STATE_RST, NULL, 0, 0,
+			       rss_map->indir_qp);
+		mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
+		mlx4_qp_free(mdev->dev, rss_map->indir_qp);
+		kfree(rss_map->indir_qp);
+		rss_map->indir_qp = NULL;
+	}
 
 	for (i = 0; i < priv->rx_ring_num; i++) {
 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
diff --git a/drivers/net/ethernet/mellanox/mlx4/main.c b/drivers/net/ethernet/mellanox/mlx4/main.c
index ccae3c6593c4..457e070bca46 100644
--- a/drivers/net/ethernet/mellanox/mlx4/main.c
+++ b/drivers/net/ethernet/mellanox/mlx4/main.c
@@ -2356,8 +2356,8 @@ static int mlx4_init_hca(struct mlx4_dev *dev)
 					MLX4_A0_STEERING_TABLE_SIZE;
 			}
 
-			mlx4_dbg(dev, "DMFS high rate steer mode is: %s\n",
-				 dmfs_high_rate_steering_mode_str(
+			mlx4_info(dev, "DMFS high rate steer mode is: %s\n",
+				  dmfs_high_rate_steering_mode_str(
 					dev->caps.dmfs_high_steer_mode));
 		}
 	} else {
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 08b2906a98af..41f4f8f9f300 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -431,7 +431,7 @@ struct mlx4_en_rss_map {
 	int base_qpn;
 	struct mlx4_qp qps[MAX_RX_RINGS];
 	enum mlx4_qp_state state[MAX_RX_RINGS];
-	struct mlx4_qp indir_qp;
+	struct mlx4_qp *indir_qp;
 	enum mlx4_qp_state indir_state;
 };
 
-- 
1.8.3.1

[PATCH net-next 03/10] net/mlx4_en: Improve receive data-path

[Tariq Toukan]

Several small performance improvements in RX datapath,
including:
- Compiler branch predictor hints.
- Replace a multiplication with a shift operation.
- Minimize variables scope.
- Write-prefetch for packet header.
- Avoid trinary-operator ("?") when value can be preset in a matching
  branch.
- Save a branch by updating RX ring doorbell within
  mlx4_en_refill_rx_buffers(), which now returns void.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON
(enable by ethtool -L <interface> rx 1).

XDP_DROP packet rate:
Same (28.1 Mpps), lower CPU utilization (from ~100% to ~92%).

Drop packets in TC:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 4.14 Mpps | 4.18 Mpps |   1% |
-------------------------------------

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 10.1 Mpps | 10.3 Mpps |   2% |
IPv6 | 10.1 Mpps | 10.3 Mpps |   2% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_rx.c | 73 ++++++++++++++++--------------
 1 file changed, 39 insertions(+), 34 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index c4edae854f1b..507c48ef2674 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
@@ -134,10 +134,11 @@ static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
 				   struct mlx4_en_rx_ring *ring, int index,
 				   gfp_t gfp)
 {
-	struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
+	struct mlx4_en_rx_desc *rx_desc = ring->buf +
+		(index << ring->log_stride);
 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
 					(index << priv->log_rx_info);
-	if (ring->page_cache.index > 0) {
+	if (likely(ring->page_cache.index > 0)) {
 		/* XDP uses a single page per frame */
 		if (!frags->page) {
 			ring->page_cache.index--;
@@ -178,6 +179,7 @@ static void mlx4_en_free_rx_desc(const struct mlx4_en_priv *priv,
 	}
 }
 
+/* Function not in fast-path */
 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
 {
 	struct mlx4_en_rx_ring *ring;
@@ -539,14 +541,14 @@ static void validate_loopback(struct mlx4_en_priv *priv, void *va)
 	priv->loopback_ok = 1;
 }
 
-static bool mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
+static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
 				      struct mlx4_en_rx_ring *ring)
 {
 	u32 missing = ring->actual_size - (ring->prod - ring->cons);
 
 	/* Try to batch allocations, but not too much. */
 	if (missing < 8)
-		return false;
+		return;
 	do {
 		if (mlx4_en_prepare_rx_desc(priv, ring,
 					    ring->prod & ring->size_mask,
@@ -554,9 +556,9 @@ static bool mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
 					    __GFP_MEMALLOC))
 			break;
 		ring->prod++;
-	} while (--missing);
+	} while (likely(--missing));
 
-	return true;
+	mlx4_en_update_rx_prod_db(ring);
 }
 
 /* When hardware doesn't strip the vlan, we need to calculate the checksum
@@ -637,21 +639,14 @@ static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
-	struct mlx4_en_dev *mdev = priv->mdev;
-	struct mlx4_cqe *cqe;
-	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
-	struct mlx4_en_rx_alloc *frags;
+	int factor = priv->cqe_factor;
+	struct mlx4_en_rx_ring *ring;
 	struct bpf_prog *xdp_prog;
+	int cq_ring = cq->ring;
 	int doorbell_pending;
-	struct sk_buff *skb;
-	int index;
-	int nr;
-	unsigned int length;
+	struct mlx4_cqe *cqe;
 	int polled = 0;
-	int ip_summed;
-	int factor = priv->cqe_factor;
-	u64 timestamp;
-	bool l2_tunnel;
+	int index;
 
 	if (unlikely(!priv->port_up))
 		return 0;
@@ -659,6 +654,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	if (unlikely(budget <= 0))
 		return polled;
 
+	ring = priv->rx_ring[cq_ring];
+
 	/* Protect accesses to: ring->xdp_prog, priv->mac_hash list */
 	rcu_read_lock();
 	xdp_prog = rcu_dereference(ring->xdp_prog);
@@ -673,10 +670,17 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	/* Process all completed CQEs */
 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
 		    cq->mcq.cons_index & cq->size)) {
+		struct mlx4_en_rx_alloc *frags;
+		enum pkt_hash_types hash_type;
+		struct sk_buff *skb;
+		unsigned int length;
+		int ip_summed;
 		void *va;
+		int nr;
 
 		frags = ring->rx_info + (index << priv->log_rx_info);
 		va = page_address(frags[0].page) + frags[0].page_offset;
+		prefetchw(va);
 		/*
 		 * make sure we read the CQE after we read the ownership bit
 		 */
@@ -768,7 +772,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 				break;
 			case XDP_TX:
 				if (likely(!mlx4_en_xmit_frame(ring, frags, dev,
-							length, cq->ring,
+							length, cq_ring,
 							&doorbell_pending))) {
 					frags[0].page = NULL;
 					goto next;
@@ -790,24 +794,27 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 		ring->packets++;
 
 		skb = napi_get_frags(&cq->napi);
-		if (!skb)
+		if (unlikely(!skb))
 			goto next;
 
 		if (unlikely(ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL)) {
-			timestamp = mlx4_en_get_cqe_ts(cqe);
-			mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
+			u64 timestamp = mlx4_en_get_cqe_ts(cqe);
+
+			mlx4_en_fill_hwtstamps(priv->mdev, skb_hwtstamps(skb),
 					       timestamp);
 		}
-		skb_record_rx_queue(skb, cq->ring);
+		skb_record_rx_queue(skb, cq_ring);
 
 		if (likely(dev->features & NETIF_F_RXCSUM)) {
 			if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
 						      MLX4_CQE_STATUS_UDP)) {
 				if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
 				    cqe->checksum == cpu_to_be16(0xffff)) {
-					ip_summed = CHECKSUM_UNNECESSARY;
-					l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
+					bool l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
 						(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
+
+					ip_summed = CHECKSUM_UNNECESSARY;
+					hash_type = PKT_HASH_TYPE_L4;
 					if (l2_tunnel)
 						skb->csum_level = 1;
 					ring->csum_ok++;
@@ -822,6 +829,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 						goto csum_none;
 					} else {
 						ip_summed = CHECKSUM_COMPLETE;
+						hash_type = PKT_HASH_TYPE_L3;
 						ring->csum_complete++;
 					}
 				} else {
@@ -831,16 +839,14 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 		} else {
 csum_none:
 			ip_summed = CHECKSUM_NONE;
+			hash_type = PKT_HASH_TYPE_L3;
 			ring->csum_none++;
 		}
 		skb->ip_summed = ip_summed;
 		if (dev->features & NETIF_F_RXHASH)
 			skb_set_hash(skb,
 				     be32_to_cpu(cqe->immed_rss_invalid),
-				     (ip_summed == CHECKSUM_UNNECESSARY) ?
-					PKT_HASH_TYPE_L4 :
-					PKT_HASH_TYPE_L3);
-
+				     hash_type);
 
 		if ((cqe->vlan_my_qpn &
 		     cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
@@ -867,13 +873,13 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 		++cq->mcq.cons_index;
 		index = (cq->mcq.cons_index) & ring->size_mask;
 		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
-		if (++polled == budget)
+		if (unlikely(++polled == budget))
 			break;
 	}
 
 	rcu_read_unlock();
 
-	if (polled) {
+	if (likely(polled)) {
 		if (doorbell_pending)
 			mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq->ring]);
 
@@ -883,8 +889,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	}
 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
 
-	if (mlx4_en_refill_rx_buffers(priv, ring))
-		mlx4_en_update_rx_prod_db(ring);
+	mlx4_en_refill_rx_buffers(priv, ring);
 
 	return polled;
 }
@@ -936,7 +941,7 @@ int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
 			done--;
 	}
 	/* Done for now */
-	if (napi_complete_done(napi, done))
+	if (likely(napi_complete_done(napi, done)))
 		mlx4_en_arm_cq(priv, cq);
 	return done;
 }
-- 
1.8.3.1

[PATCH net-next 04/10] net/mlx4_en: Improve transmit CQ polling

[Tariq Toukan]

Several small performance improvements in TX CQ polling,
including:
- Compiler branch predictor hints.
- Minimize variables scope.
- More proper check of cq type.
- Use boolean instead of int for a binary indication.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Packet-rate tests for both regular stack and XDP use cases:
No noticeable gain, no degradation.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 37386abea54c..2d5e0da1de2f 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -402,8 +402,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	struct mlx4_cq *mcq = &cq->mcq;
 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->type][cq->ring];
 	struct mlx4_cqe *cqe;
-	u16 index;
-	u16 new_index, ring_index, stamp_index;
+	u16 index, ring_index, stamp_index;
 	u32 txbbs_skipped = 0;
 	u32 txbbs_stamp = 0;
 	u32 cons_index = mcq->cons_index;
@@ -418,7 +417,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	u32 last_nr_txbb;
 	u32 ring_cons;
 
-	if (!priv->port_up)
+	if (unlikely(!priv->port_up))
 		return true;
 
 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
@@ -433,6 +432,8 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	/* Process all completed CQEs */
 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
 			cons_index & size) && (done < budget)) {
+		u16 new_index;
+
 		/*
 		 * make sure we read the CQE after we read the
 		 * ownership bit
@@ -479,7 +480,6 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
 	}
 
-
 	/*
 	 * To prevent CQ overflow we first update CQ consumer and only then
 	 * the ring consumer.
@@ -492,7 +492,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
 	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
 
-	if (ring->free_tx_desc == mlx4_en_recycle_tx_desc)
+	if (cq->type == TX_XDP)
 		return done < budget;
 
 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
@@ -504,6 +504,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 		netif_tx_wake_queue(ring->tx_queue);
 		ring->wake_queue++;
 	}
+
 	return done < budget;
 }
 
@@ -524,7 +525,7 @@ int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
 	struct net_device *dev = cq->dev;
 	struct mlx4_en_priv *priv = netdev_priv(dev);
-	int clean_complete;
+	bool clean_complete;
 
 	clean_complete = mlx4_en_process_tx_cq(dev, cq, budget);
 	if (!clean_complete)
-- 
1.8.3.1

[PATCH net-next 05/10] net/mlx4_en: Improve stack xmit function

[Tariq Toukan]

Several small code and performance improvements in stack TX datapath,
including:
- Compiler branch predictor hints.
- Minimize variables scope.
- Move tx_info non-inline flow handling to a separate function.
- Calculate data_offset in compile time rather than in runtime
  (for !lso_header_size branch).
- Avoid trinary-operator ("?") when value can be preset in a matching
  branch.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 151 +++++++++++++++++------------
 1 file changed, 87 insertions(+), 64 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 2d5e0da1de2f..58f4b322587b 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -774,37 +774,101 @@ static void mlx4_en_tx_write_desc(struct mlx4_en_tx_ring *ring,
 	}
 }
 
+static bool mlx4_en_build_dma_wqe(struct mlx4_en_priv *priv,
+				  struct skb_shared_info *shinfo,
+				  struct mlx4_wqe_data_seg *data,
+				  struct sk_buff *skb,
+				  int lso_header_size,
+				  __be32 mr_key,
+				  struct mlx4_en_tx_info *tx_info)
+{
+	struct device *ddev = priv->ddev;
+	dma_addr_t dma = 0;
+	u32 byte_count = 0;
+	int i_frag;
+
+	/* Map fragments if any */
+	for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
+		const struct skb_frag_struct *frag;
+
+		frag = &shinfo->frags[i_frag];
+		byte_count = skb_frag_size(frag);
+		dma = skb_frag_dma_map(ddev, frag,
+				       0, byte_count,
+				       DMA_TO_DEVICE);
+		if (dma_mapping_error(ddev, dma))
+			goto tx_drop_unmap;
+
+		data->addr = cpu_to_be64(dma);
+		data->lkey = mr_key;
+		dma_wmb();
+		data->byte_count = cpu_to_be32(byte_count);
+		--data;
+	}
+
+	/* Map linear part if needed */
+	if (tx_info->linear) {
+		byte_count = skb_headlen(skb) - lso_header_size;
+
+		dma = dma_map_single(ddev, skb->data +
+				     lso_header_size, byte_count,
+				     PCI_DMA_TODEVICE);
+		if (dma_mapping_error(ddev, dma))
+			goto tx_drop_unmap;
+
+		data->addr = cpu_to_be64(dma);
+		data->lkey = mr_key;
+		dma_wmb();
+		data->byte_count = cpu_to_be32(byte_count);
+	}
+	/* tx completion can avoid cache line miss for common cases */
+	tx_info->map0_dma = dma;
+	tx_info->map0_byte_count = byte_count;
+
+	return true;
+
+tx_drop_unmap:
+	en_err(priv, "DMA mapping error\n");
+
+	while (++i_frag < shinfo->nr_frags) {
+		++data;
+		dma_unmap_page(ddev, (dma_addr_t)be64_to_cpu(data->addr),
+			       be32_to_cpu(data->byte_count),
+			       PCI_DMA_TODEVICE);
+	}
+
+	return false;
+}
+
 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct skb_shared_info *shinfo = skb_shinfo(skb);
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	union mlx4_wqe_qpn_vlan	qpn_vlan = {};
-	struct device *ddev = priv->ddev;
 	struct mlx4_en_tx_ring *ring;
 	struct mlx4_en_tx_desc *tx_desc;
 	struct mlx4_wqe_data_seg *data;
 	struct mlx4_en_tx_info *tx_info;
-	int tx_ind = 0;
+	int tx_ind;
 	int nr_txbb;
 	int desc_size;
 	int real_size;
 	u32 index, bf_index;
 	__be32 op_own;
-	u16 vlan_proto = 0;
-	int i_frag;
 	int lso_header_size;
 	void *fragptr = NULL;
 	bool bounce = false;
 	bool send_doorbell;
 	bool stop_queue;
 	bool inline_ok;
+	u8 data_offset;
 	u32 ring_cons;
 	bool bf_ok;
 
 	tx_ind = skb_get_queue_mapping(skb);
 	ring = priv->tx_ring[TX][tx_ind];
 
-	if (!priv->port_up)
+	if (unlikely(!priv->port_up))
 		goto tx_drop;
 
 	/* fetch ring->cons far ahead before needing it to avoid stall */
@@ -826,6 +890,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	bf_ok = ring->bf_enabled;
 	if (skb_vlan_tag_present(skb)) {
+		u16 vlan_proto;
+
 		qpn_vlan.vlan_tag = cpu_to_be16(skb_vlan_tag_get(skb));
 		vlan_proto = be16_to_cpu(skb->vlan_proto);
 		if (vlan_proto == ETH_P_8021AD)
@@ -862,64 +928,31 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	tx_info->skb = skb;
 	tx_info->nr_txbb = nr_txbb;
 
-	data = &tx_desc->data;
-	if (lso_header_size)
-		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
-						      DS_SIZE));
+	if (!lso_header_size) {
+		data = &tx_desc->data;
+		data_offset = offsetof(struct mlx4_en_tx_desc, data);
+	} else {
+		int lso_align = ALIGN(lso_header_size + 4, DS_SIZE);
+
+		data = (void *)&tx_desc->lso + lso_align;
+		data_offset = offsetof(struct mlx4_en_tx_desc, lso) + lso_align;
+	}
 
 	/* valid only for none inline segments */
-	tx_info->data_offset = (void *)data - (void *)tx_desc;
+	tx_info->data_offset = data_offset;
 
 	tx_info->inl = inline_ok;
 
-	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
-			   !inline_ok) ? 1 : 0;
+	tx_info->linear = lso_header_size < skb_headlen(skb) && !inline_ok;
 
 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
 	data += tx_info->nr_maps - 1;
 
-	if (!tx_info->inl) {
-		dma_addr_t dma = 0;
-		u32 byte_count = 0;
-
-		/* Map fragments if any */
-		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
-			const struct skb_frag_struct *frag;
-
-			frag = &shinfo->frags[i_frag];
-			byte_count = skb_frag_size(frag);
-			dma = skb_frag_dma_map(ddev, frag,
-					       0, byte_count,
-					       DMA_TO_DEVICE);
-			if (dma_mapping_error(ddev, dma))
-				goto tx_drop_unmap;
-
-			data->addr = cpu_to_be64(dma);
-			data->lkey = ring->mr_key;
-			dma_wmb();
-			data->byte_count = cpu_to_be32(byte_count);
-			--data;
-		}
-
-		/* Map linear part if needed */
-		if (tx_info->linear) {
-			byte_count = skb_headlen(skb) - lso_header_size;
-
-			dma = dma_map_single(ddev, skb->data +
-					     lso_header_size, byte_count,
-					     PCI_DMA_TODEVICE);
-			if (dma_mapping_error(ddev, dma))
-				goto tx_drop_unmap;
-
-			data->addr = cpu_to_be64(dma);
-			data->lkey = ring->mr_key;
-			dma_wmb();
-			data->byte_count = cpu_to_be32(byte_count);
-		}
-		/* tx completion can avoid cache line miss for common cases */
-		tx_info->map0_dma = dma;
-		tx_info->map0_byte_count = byte_count;
-	}
+	if (!tx_info->inl)
+		if (!mlx4_en_build_dma_wqe(priv, shinfo, data, skb,
+					   lso_header_size, ring->mr_key,
+					   tx_info))
+			goto tx_drop_count;
 
 	/*
 	 * For timestamping add flag to skb_shinfo and
@@ -1055,16 +1088,6 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	}
 	return NETDEV_TX_OK;
 
-tx_drop_unmap:
-	en_err(priv, "DMA mapping error\n");
-
-	while (++i_frag < shinfo->nr_frags) {
-		++data;
-		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
-			       be32_to_cpu(data->byte_count),
-			       PCI_DMA_TODEVICE);
-	}
-
 tx_drop_count:
 	ring->tx_dropped++;
 tx_drop:
-- 
1.8.3.1

[PATCH net-next 06/10] net/mlx4_en: Improve XDP xmit function

[Tariq Toukan]

Several performance improvements in XDP TX datapath,
including:
- Ring a single doorbell for XDP TX ring per NAPI budget,
  instead of doing it per a lower threshold (was 8).
  This includes removing the flow of immediate doorbell ringing
  in case of a full TX ring.
- Compiler branch predictor hints.
- Calculate values in compile time rather than in runtime.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON.

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 10.3 Mpps | 12.0 Mpps |  17% |
IPv6 | 10.3 Mpps | 12.0 Mpps |  17% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_rx.c   |  2 +-
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 59 +++++++++-------------------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  3 +-
 3 files changed, 21 insertions(+), 43 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index 507c48ef2674..747e4d7d7693 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
@@ -643,7 +643,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	struct mlx4_en_rx_ring *ring;
 	struct bpf_prog *xdp_prog;
 	int cq_ring = cq->ring;
-	int doorbell_pending;
+	bool doorbell_pending;
 	struct mlx4_cqe *cqe;
 	int polled = 0;
 	int index;
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 58f4b322587b..01bb43879221 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -1095,51 +1095,40 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	return NETDEV_TX_OK;
 }
 
+#define MLX4_EN_XDP_TX_NRTXBB  1
+#define MLX4_EN_XDP_TX_REAL_SZ (((CTRL_SIZE + MLX4_EN_XDP_TX_NRTXBB * DS_SIZE) \
+				 / 16) & 0x3f)
+
 netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 			       struct mlx4_en_rx_alloc *frame,
 			       struct net_device *dev, unsigned int length,
-			       int tx_ind, int *doorbell_pending)
+			       int tx_ind, bool *doorbell_pending)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	union mlx4_wqe_qpn_vlan	qpn_vlan = {};
-	struct mlx4_en_tx_ring *ring;
 	struct mlx4_en_tx_desc *tx_desc;
-	struct mlx4_wqe_data_seg *data;
 	struct mlx4_en_tx_info *tx_info;
-	int index, bf_index;
-	bool send_doorbell;
-	int nr_txbb = 1;
-	bool stop_queue;
+	struct mlx4_wqe_data_seg *data;
+	struct mlx4_en_tx_ring *ring;
 	dma_addr_t dma;
-	int real_size;
 	__be32 op_own;
-	u32 ring_cons;
-	bool bf_ok;
+	int index;
 
-	BUILD_BUG_ON_MSG(ALIGN(CTRL_SIZE + DS_SIZE, TXBB_SIZE) != TXBB_SIZE,
-			 "mlx4_en_xmit_frame requires minimum size tx desc");
+	if (unlikely(!priv->port_up))
+		goto tx_drop;
 
 	ring = priv->tx_ring[TX_XDP][tx_ind];
 
-	if (!priv->port_up)
-		goto tx_drop;
-
-	if (mlx4_en_is_tx_ring_full(ring))
+	if (unlikely(mlx4_en_is_tx_ring_full(ring)))
 		goto tx_drop_count;
 
-	/* fetch ring->cons far ahead before needing it to avoid stall */
-	ring_cons = READ_ONCE(ring->cons);
-
 	index = ring->prod & ring->size_mask;
 	tx_info = &ring->tx_info[index];
 
-	bf_ok = ring->bf_enabled;
-
 	/* Track current inflight packets for performance analysis */
 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
-			 (u32)(ring->prod - ring_cons - 1));
+			 (u32)(ring->prod - READ_ONCE(ring->cons) - 1));
 
-	bf_index = ring->prod;
 	tx_desc = ring->buf + index * TXBB_SIZE;
 	data = &tx_desc->data;
 
@@ -1149,9 +1138,9 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 	frame->page = NULL;
 	tx_info->map0_dma = dma;
 	tx_info->map0_byte_count = PAGE_SIZE;
-	tx_info->nr_txbb = nr_txbb;
+	tx_info->nr_txbb = MLX4_EN_XDP_TX_NRTXBB;
 	tx_info->nr_bytes = max_t(unsigned int, length, ETH_ZLEN);
-	tx_info->data_offset = (void *)data - (void *)tx_desc;
+	tx_info->data_offset = offsetof(struct mlx4_en_tx_desc, data);
 	tx_info->ts_requested = 0;
 	tx_info->nr_maps = 1;
 	tx_info->linear = 1;
@@ -1175,23 +1164,13 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 	rx_ring->xdp_tx++;
 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, length);
 
-	ring->prod += nr_txbb;
+	ring->prod += MLX4_EN_XDP_TX_NRTXBB;
 
-	stop_queue = mlx4_en_is_tx_ring_full(ring);
-	send_doorbell = stop_queue ||
-				*doorbell_pending > MLX4_EN_DOORBELL_BUDGET;
-	bf_ok &= send_doorbell;
+	qpn_vlan.fence_size = MLX4_EN_XDP_TX_REAL_SZ;
 
-	real_size = ((CTRL_SIZE + nr_txbb * DS_SIZE) / 16) & 0x3f;
-
-	if (bf_ok)
-		qpn_vlan.bf_qpn = ring->doorbell_qpn | cpu_to_be32(real_size);
-	else
-		qpn_vlan.fence_size = real_size;
-
-	mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, TXBB_SIZE, bf_index,
-			      op_own, bf_ok, send_doorbell);
-	*doorbell_pending = send_doorbell ? 0 : *doorbell_pending + 1;
+	mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, TXBB_SIZE, 0,
+			      op_own, false, false);
+	*doorbell_pending = true;
 
 	return NETDEV_TX_OK;
 
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 41f4f8f9f300..c52edb717add 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -121,7 +121,6 @@
 					 MLX4_EN_NUM_UP)
 
 #define MLX4_EN_DEFAULT_TX_WORK		256
-#define MLX4_EN_DOORBELL_BUDGET		8
 
 /* Target number of packets to coalesce with interrupt moderation */
 #define MLX4_EN_RX_COAL_TARGET	44
@@ -689,7 +688,7 @@ u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
 netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 			       struct mlx4_en_rx_alloc *frame,
 			       struct net_device *dev, unsigned int length,
-			       int tx_ind, int *doorbell_pending);
+			       int tx_ind, bool *doorbell_pending);
 void mlx4_en_xmit_doorbell(struct mlx4_en_tx_ring *ring);
 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
 			struct mlx4_en_rx_alloc *frame);
-- 
1.8.3.1

[PATCH net-next 07/10] net/mlx4_en: Poll XDP TX completion queue in RX NAPI

[Tariq Toukan]

Instead of having their own NAPIs, XDP TX completion queues get
polled within the corresponding RX NAPI.
This prevents any possible race on TX ring prod/cons indices,
between the context that issues the transmits (RX NAPI) and the
context that handles the completions (was previously done in
a separate NAPI).

This also improves performance, as it decreases the number
of NAPIs running on a CPU, saving the overhead of syncing
and switching between the contexts.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON.

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 12.0 Mpps | 13.8 Mpps |  15% |
IPv6 | 12.0 Mpps | 13.8 Mpps |  15% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_cq.c     | 25 ++++++++++++++++++-------
 drivers/net/ethernet/mellanox/mlx4/en_netdev.c |  8 +++++---
 drivers/net/ethernet/mellanox/mlx4/en_rx.c     | 22 +++++++++++++++++++---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c     |  5 +++--
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h   |  7 ++++++-
 5 files changed, 51 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_cq.c b/drivers/net/ethernet/mellanox/mlx4/en_cq.c
index 09dd3776db76..85fe17e4dcfb 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_cq.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_cq.c
@@ -146,16 +146,25 @@ int mlx4_en_activate_cq(struct mlx4_en_priv *priv, struct mlx4_en_cq *cq,
 	if (err)
 		goto free_eq;
 
-	cq->mcq.comp  = cq->type != RX ? mlx4_en_tx_irq : mlx4_en_rx_irq;
 	cq->mcq.event = mlx4_en_cq_event;
 
-	if (cq->type != RX)
+	switch (cq->type) {
+	case TX:
+		cq->mcq.comp = mlx4_en_tx_irq;
 		netif_tx_napi_add(cq->dev, &cq->napi, mlx4_en_poll_tx_cq,
 				  NAPI_POLL_WEIGHT);
-	else
+		napi_enable(&cq->napi);
+		break;
+	case RX:
+		cq->mcq.comp = mlx4_en_rx_irq;
 		netif_napi_add(cq->dev, &cq->napi, mlx4_en_poll_rx_cq, 64);
-
-	napi_enable(&cq->napi);
+		napi_enable(&cq->napi);
+		break;
+	case TX_XDP:
+		/* nothing regarding napi, it's shared with rx ring */
+		cq->xdp_busy = false;
+		break;
+	}
 
 	return 0;
 
@@ -184,8 +193,10 @@ void mlx4_en_destroy_cq(struct mlx4_en_priv *priv, struct mlx4_en_cq **pcq)
 
 void mlx4_en_deactivate_cq(struct mlx4_en_priv *priv, struct mlx4_en_cq *cq)
 {
-	napi_disable(&cq->napi);
-	netif_napi_del(&cq->napi);
+	if (cq->type != TX_XDP) {
+		napi_disable(&cq->napi);
+		netif_napi_del(&cq->napi);
+	}
 
 	mlx4_cq_free(priv->mdev->dev, &cq->mcq);
 }
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
index 51ce111b719e..99c02bb4f302 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
@@ -1679,13 +1679,15 @@ int mlx4_en_start_port(struct net_device *dev)
 			if (t != TX_XDP) {
 				tx_ring->tx_queue = netdev_get_tx_queue(dev, i);
 				tx_ring->recycle_ring = NULL;
+
+				/* Arm CQ for TX completions */
+				mlx4_en_arm_cq(priv, cq);
+
 			} else {
 				mlx4_en_init_recycle_ring(priv, i);
+				/* XDP TX CQ should never be armed */
 			}
 
-			/* Arm CQ for TX completions */
-			mlx4_en_arm_cq(priv, cq);
-
 			/* Set initial ownership of all Tx TXBBs to SW (1) */
 			for (j = 0; j < tx_ring->buf_size; j += STAMP_STRIDE)
 				*((u32 *)(tx_ring->buf + j)) = 0xffffffff;
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index 747e4d7d7693..e5fb89505a13 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
@@ -880,8 +880,10 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	rcu_read_unlock();
 
 	if (likely(polled)) {
-		if (doorbell_pending)
-			mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq->ring]);
+		if (doorbell_pending) {
+			priv->tx_cq[TX_XDP][cq_ring]->xdp_busy = true;
+			mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq_ring]);
+		}
 
 		mlx4_cq_set_ci(&cq->mcq);
 		wmb(); /* ensure HW sees CQ consumer before we post new buffers */
@@ -912,16 +914,30 @@ int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
 	struct net_device *dev = cq->dev;
 	struct mlx4_en_priv *priv = netdev_priv(dev);
+	struct mlx4_en_cq *xdp_tx_cq = NULL;
+	bool clean_complete = true;
 	int done;
 
+	if (priv->tx_ring_num[TX_XDP]) {
+		xdp_tx_cq = priv->tx_cq[TX_XDP][cq->ring];
+		if (xdp_tx_cq->xdp_busy) {
+			clean_complete = mlx4_en_process_tx_cq(dev, xdp_tx_cq,
+							       budget);
+			xdp_tx_cq->xdp_busy = !clean_complete;
+		}
+	}
+
 	done = mlx4_en_process_rx_cq(dev, cq, budget);
 
 	/* If we used up all the quota - we're probably not done yet... */
-	if (done == budget) {
+	if (done == budget || !clean_complete) {
 		const struct cpumask *aff;
 		struct irq_data *idata;
 		int cpu_curr;
 
+		/* in case we got here because of !clean_complete */
+		done = budget;
+
 		INC_PERF_COUNTER(priv->pstats.napi_quota);
 
 		cpu_curr = smp_processor_id();
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 01bb43879221..500442c60342 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -395,8 +395,8 @@ int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
 	return cnt;
 }
 
-static bool mlx4_en_process_tx_cq(struct net_device *dev,
-				  struct mlx4_en_cq *cq, int napi_budget)
+bool mlx4_en_process_tx_cq(struct net_device *dev,
+			   struct mlx4_en_cq *cq, int napi_budget)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	struct mlx4_cq *mcq = &cq->mcq;
@@ -1176,6 +1176,7 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 
 tx_drop_count:
 	rx_ring->xdp_tx_full++;
+	*doorbell_pending = true;
 tx_drop:
 	return NETDEV_TX_BUSY;
 }
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index c52edb717add..bde3bc869b70 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -358,7 +358,10 @@ struct mlx4_en_cq {
 	struct mlx4_hwq_resources wqres;
 	int                     ring;
 	struct net_device      *dev;
-	struct napi_struct	napi;
+	union {
+		struct napi_struct napi;
+		bool               xdp_busy;
+	};
 	int size;
 	int buf_size;
 	int vector;
@@ -720,6 +723,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev,
 			  int budget);
 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget);
 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget);
+bool mlx4_en_process_tx_cq(struct net_device *dev,
+			   struct mlx4_en_cq *cq, int napi_budget);
 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 			 struct mlx4_en_tx_ring *ring,
 			 int index, u64 timestamp,
-- 
1.8.3.1

[PATCH net-next 08/10] net/mlx4_en: Increase default TX ring size

[Tariq Toukan]

Increase the default TX ring size (from 512 to 1024) to match
the RX ring size.
This gives the XDP TX ring a better chance to keep up with the
rate of its RX ring in case of a high load of XDP_TX actions.

Tested:
Ethtool counter rx_xdp_tx_full used to increase, after applying this
patch it stopped.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index bde3bc869b70..7f61b5829709 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -115,8 +115,8 @@
 #define MLX4_EN_MIN_TX_RING_P_UP	1
 #define MLX4_EN_MAX_TX_RING_P_UP	32
 #define MLX4_EN_NUM_UP			8
-#define MLX4_EN_DEF_TX_RING_SIZE	512
 #define MLX4_EN_DEF_RX_RING_SIZE  	1024
+#define MLX4_EN_DEF_TX_RING_SIZE	MLX4_EN_DEF_RX_RING_SIZE
 #define MAX_TX_RINGS			(MLX4_EN_MAX_TX_RING_P_UP * \
 					 MLX4_EN_NUM_UP)
 
-- 
1.8.3.1

[PATCH net-next 09/10] net/mlx4_en: Replace TXBB_SIZE multiplications with shift operations

[Tariq Toukan]

Define LOG_TXBB_SIZE, log of TXBB_SIZE, and use it with a shift
operation instead of a multiplication with TXBB_SIZE.
Operations are equivalent as TXBB_SIZE is a power of two.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 26 ++++++++++++++------------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  3 ++-
 2 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 500442c60342..efc4411b1e44 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -234,23 +234,24 @@ static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
 			      u8 owner)
 {
 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
-	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
+	struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
 	void *end = ring->buf + ring->buf_size;
 	__be32 *ptr = (__be32 *)tx_desc;
 	int i;
 
 	/* Optimize the common case when there are no wraparounds */
-	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
+	if (likely((void *)tx_desc +
+		   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
 		/* Stamp the freed descriptor */
-		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
+		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
 		     i += STAMP_STRIDE) {
 			*ptr = stamp;
 			ptr += STAMP_DWORDS;
 		}
 	} else {
 		/* Stamp the freed descriptor */
-		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
+		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
 		     i += STAMP_STRIDE) {
 			*ptr = stamp;
 			ptr += STAMP_DWORDS;
@@ -269,7 +270,7 @@ u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 			 int napi_mode)
 {
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
-	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
+	struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
 	void *end = ring->buf + ring->buf_size;
 	struct sk_buff *skb = tx_info->skb;
@@ -289,7 +290,8 @@ u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 	}
 
 	/* Optimize the common case when there are no wraparounds */
-	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
+	if (likely((void *)tx_desc +
+		   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
 		if (!tx_info->inl) {
 			if (tx_info->linear)
 				dma_unmap_single(priv->ddev,
@@ -542,7 +544,7 @@ static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 						      u32 index,
 						      unsigned int desc_size)
 {
-	u32 copy = (ring->size - index) * TXBB_SIZE;
+	u32 copy = (ring->size - index) << LOG_TXBB_SIZE;
 	int i;
 
 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
@@ -557,12 +559,12 @@ static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 		if ((i & (TXBB_SIZE - 1)) == 0)
 			wmb();
 
-		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
+		*((u32 *)(ring->buf + (index << LOG_TXBB_SIZE) + i)) =
 			*((u32 *) (ring->bounce_buf + i));
 	}
 
 	/* Return real descriptor location */
-	return ring->buf + index * TXBB_SIZE;
+	return ring->buf + (index << LOG_TXBB_SIZE);
 }
 
 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
@@ -881,7 +883,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	/* Align descriptor to TXBB size */
 	desc_size = ALIGN(real_size, TXBB_SIZE);
-	nr_txbb = desc_size / TXBB_SIZE;
+	nr_txbb = desc_size >> LOG_TXBB_SIZE;
 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
 		if (netif_msg_tx_err(priv))
 			en_warn(priv, "Oversized header or SG list\n");
@@ -916,7 +918,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	/* See if we have enough space for whole descriptor TXBB for setting
 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
 	if (likely(index + nr_txbb <= ring->size))
-		tx_desc = ring->buf + index * TXBB_SIZE;
+		tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
 	else {
 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
 		bounce = true;
@@ -1129,7 +1131,7 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
 			 (u32)(ring->prod - READ_ONCE(ring->cons) - 1));
 
-	tx_desc = ring->buf + index * TXBB_SIZE;
+	tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
 	data = &tx_desc->data;
 
 	dma = frame->dma;
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 7f61b5829709..963b77d51b48 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -72,7 +72,8 @@
 #define DEF_RX_RINGS		16
 #define MAX_RX_RINGS		128
 #define MIN_RX_RINGS		4
-#define TXBB_SIZE		64
+#define LOG_TXBB_SIZE		6
+#define TXBB_SIZE		BIT(LOG_TXBB_SIZE)
 #define HEADROOM		(2048 / TXBB_SIZE + 1)
 #define STAMP_STRIDE		64
 #define STAMP_DWORDS		(STAMP_STRIDE / 4)
-- 
1.8.3.1

[PATCH net-next 10/10] net/mlx4_en: Refactor mlx4_en_free_tx_desc

[Tariq Toukan]

Some code re-ordering, functionally equivalent.

- The !tx_info->inl check is evaluated anyway in both flows
  (common case/end case). Run it first, this might finish
  the flows earlier.
- dma_unmap calls are identical in both flows, get it out
  of the if block into the common area.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 45 +++++++++++-------------------
 1 file changed, 16 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index efc4411b1e44..7d69d939ee2d 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -289,20 +289,20 @@ u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 		skb_tstamp_tx(skb, &hwts);
 	}
 
-	/* Optimize the common case when there are no wraparounds */
-	if (likely((void *)tx_desc +
-		   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
-		if (!tx_info->inl) {
-			if (tx_info->linear)
-				dma_unmap_single(priv->ddev,
-						tx_info->map0_dma,
-						tx_info->map0_byte_count,
-						PCI_DMA_TODEVICE);
-			else
-				dma_unmap_page(priv->ddev,
-					       tx_info->map0_dma,
-					       tx_info->map0_byte_count,
-					       PCI_DMA_TODEVICE);
+	if (!tx_info->inl) {
+		if (tx_info->linear)
+			dma_unmap_single(priv->ddev,
+					 tx_info->map0_dma,
+					 tx_info->map0_byte_count,
+					 PCI_DMA_TODEVICE);
+		else
+			dma_unmap_page(priv->ddev,
+				       tx_info->map0_dma,
+				       tx_info->map0_byte_count,
+				       PCI_DMA_TODEVICE);
+		/* Optimize the common case when there are no wraparounds */
+		if (likely((void *)tx_desc +
+			   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
 			for (i = 1; i < nr_maps; i++) {
 				data++;
 				dma_unmap_page(priv->ddev,
@@ -310,23 +310,10 @@ u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 					be32_to_cpu(data->byte_count),
 					PCI_DMA_TODEVICE);
 			}
-		}
-	} else {
-		if (!tx_info->inl) {
-			if ((void *) data >= end) {
+		} else {
+			if ((void *)data >= end)
 				data = ring->buf + ((void *)data - end);
-			}
 
-			if (tx_info->linear)
-				dma_unmap_single(priv->ddev,
-						tx_info->map0_dma,
-						tx_info->map0_byte_count,
-						PCI_DMA_TODEVICE);
-			else
-				dma_unmap_page(priv->ddev,
-					       tx_info->map0_dma,
-					       tx_info->map0_byte_count,
-					       PCI_DMA_TODEVICE);
 			for (i = 1; i < nr_maps; i++) {
 				data++;
 				/* Check for wraparound before unmapping */
-- 
1.8.3.1

Re: [PATCH net-next 00/10] mlx4 XDP performance improvements

[David Miller]

From: Tariq Toukan <tariqt@mellanox.com>
Date: Thu, 15 Jun 2017 14:35:30 +0300

> This patchset contains data-path improvements, mainly for XDP_DROP
> and XDP_TX cases.
> 
> Main patches:
> * Patch 2 by Saeed allows enabling optimized A0 RX steering (in HW) when
>   setting a single RX ring.
>   With this configuration, HW packet-rate dramatically improves,
>   reaching 28.1 Mpps in XDP_DROP case for both IPv4 (37% gain)
>   and IPv6 (53% gain).
> * Patch 6 enhances the XDP xmit function. Among other changes, now we
>   ring one doorbell per NAPI. Patch gives 17% gain in XDP_TX case.
> * Patch 7 obsoletes the NAPI of XDP_TX completion queue and integrates its
>   poll into the respective RX NAPI. Patch gives 15% gain in XDP_TX case.
> 
> Series generated against net-next commit:
> f7aec129a356 rxrpc: Cache the congestion window setting

Series applied, thanks.

RE: [PATCH net-next 09/10] net/mlx4_en: Replace TXBB_SIZE multiplications with shift operations

[David Laight]

From: Tariq Toukan
> Sent: 15 June 2017 12:36
> Define LOG_TXBB_SIZE, log of TXBB_SIZE, and use it with a shift
> operation instead of a multiplication with TXBB_SIZE.
> Operations are equivalent as TXBB_SIZE is a power of two.
> 
> Performance tests:
> Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
> 
> Gain is too small to be measurable, no degradation sensed.
> Results are similar for IPv4 and IPv6.

I can't imagine there is any difference at all.
The compiler will use a shift for a 'multiply by a constant power of 2'.

...
If you want to save a few cycles I think the loop:
> -		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
> +		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
requires the compiler generate code to read nr_txbb every
iteration.
Caching the values might help (unless it causes a different
register spill).

	David

Re: [PATCH net-next 09/10] net/mlx4_en: Replace TXBB_SIZE multiplications with shift operations

[Tariq Toukan]

On 20/06/2017 11:45 AM, David Laight wrote:
> From: Tariq Toukan
>> Sent: 15 June 2017 12:36
>> Define LOG_TXBB_SIZE, log of TXBB_SIZE, and use it with a shift
>> operation instead of a multiplication with TXBB_SIZE.
>> Operations are equivalent as TXBB_SIZE is a power of two.
>>
>> Performance tests:
>> Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
>>
>> Gain is too small to be measurable, no degradation sensed.
>> Results are similar for IPv4 and IPv6.
> 
> I can't imagine there is any difference at all.
> The compiler will use a shift for a 'multiply by a constant power of 2'.
Yeah i guess my compiler does that, because it's a constant known at 
compile-time.

> 
> ...
> If you want to save a few cycles I think the loop:
>> -		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
>> +		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
> requires the compiler generate code to read nr_txbb every
> iteration.
> Caching the values might help (unless it causes a different
> register spill).

That sounds good!
I'll prepare and send this after testing.
I'll also look for similar cases in driver.

> 
> 	David
> 

Thank you David!