net/mlx5e: Enable CQE based moderation on TX CQ

By using CQE based moderation on TX CQ we can reduce the number of TX
interrupt rate. Besides the benefit of less interrupts, this also
allows the kernel to better utilize TSO. Since TSO has some CPU overhead,
it might not aggregate when CPU is under high stress. By reducing the
interrupt rate and the CPU utilization, we can get better aggregation
and better overall throughput.
The feature is enabled by default and has a private flag in ethtool
for control.

Throughput, interrupt rate and TSO utilization improvements:
(ConnectX-4Lx 40GbE, unidirectional, 1/16 TCP streams, 64B packets)
---------------------------------------------------------
Metric   | Streams | CQE Based | EQE Based | improvement
---------------------------------------------------------
BW       |    1    |  2.4Gb/s  | 2.15Gb/s  |  +11.6%
IR       |    1    |  27Kips   | 50.6Kips  |  -46.7%
TSO Util |    1    |  74.6%    | 71%       |  +5%
BW       |    16   |  29Gb/s   | 25.85Gb/s |  +12.2%
IR       |    16   |  482Kips  | 745Kips   |  -35.3%
TSO Util |    16   |  69.1%    | 49%       |  +41.1%

*BW = Bandwidth, IR = Interrupt rate, ips = interrupt per second.
TSO Util = bytes in TSO sessions / all bytes transferred

Signed-off-by: Tal Gilboa <talgi@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>

drivers/net/ethernet/mellanox/mlx5/core/en.h

@@ -106,6 +106,7 @@
 #define MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE 0x3
 #define MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS      0x20
 #define MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC      0x10
+#define MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE 0x10
 #define MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS      0x20
 #define MLX5E_PARAMS_DEFAULT_MIN_RX_WQES                0x80
 #define MLX5E_PARAMS_DEFAULT_MIN_RX_WQES_MPW            0x2
@@ -198,12 +199,14 @@ extern const char mlx5e_self_tests[][ETH_GSTRING_LEN];
 
 static const char mlx5e_priv_flags[][ETH_GSTRING_LEN] = {
 	"rx_cqe_moder",
+	"tx_cqe_moder",
 	"rx_cqe_compress",
 };
 
 enum mlx5e_priv_flag {
 	MLX5E_PFLAG_RX_CQE_BASED_MODER = (1 << 0),
-	MLX5E_PFLAG_RX_CQE_COMPRESS = (1 << 1),
+	MLX5E_PFLAG_TX_CQE_BASED_MODER = (1 << 1),
+	MLX5E_PFLAG_RX_CQE_COMPRESS = (1 << 2),
 };
 
 #define MLX5E_SET_PFLAG(params, pflag, enable)			\
@@ -223,6 +226,7 @@ enum mlx5e_priv_flag {
 struct mlx5e_cq_moder {
 	u16 usec;
 	u16 pkts;
+	u8 cq_period_mode;
 };
 
 struct mlx5e_params {
@@ -234,7 +238,6 @@ struct mlx5e_params {
 	u8  log_rq_size;
 	u16 num_channels;
 	u8  num_tc;
-	u8  rx_cq_period_mode;
 	bool rx_cqe_compress_def;
 	struct mlx5e_cq_moder rx_cq_moderation;
 	struct mlx5e_cq_moder tx_cq_moderation;
@@ -926,6 +929,8 @@ void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
 				   int num_channels);
 int mlx5e_get_max_linkspeed(struct mlx5_core_dev *mdev, u32 *speed);
 
+void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params,
+				 u8 cq_period_mode);
 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params,
 				 u8 cq_period_mode);
 void mlx5e_set_rq_type_params(struct mlx5_core_dev *mdev,

drivers/net/ethernet/mellanox/mlx5/core/en_ethtool.c

@@ -1454,29 +1454,36 @@ static int mlx5e_get_module_eeprom(struct net_device *netdev,
 
 typedef int (*mlx5e_pflag_handler)(struct net_device *netdev, bool enable);
 
-static int set_pflag_rx_cqe_based_moder(struct net_device *netdev, bool enable)
+static int set_pflag_cqe_based_moder(struct net_device *netdev, bool enable,
+				     bool is_rx_cq)
 {
 	struct mlx5e_priv *priv = netdev_priv(netdev);
 	struct mlx5_core_dev *mdev = priv->mdev;
 	struct mlx5e_channels new_channels = {};
-	bool rx_mode_changed;
-	u8 rx_cq_period_mode;
+	bool mode_changed;
+	u8 cq_period_mode, current_cq_period_mode;
 	int err = 0;
 
-	rx_cq_period_mode = enable ?
+	cq_period_mode = enable ?
 		MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
 		MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
-	rx_mode_changed = rx_cq_period_mode != priv->channels.params.rx_cq_period_mode;
+	current_cq_period_mode = is_rx_cq ?
+		priv->channels.params.rx_cq_moderation.cq_period_mode :
+		priv->channels.params.tx_cq_moderation.cq_period_mode;
+	mode_changed = cq_period_mode != current_cq_period_mode;
 
-	if (rx_cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE &&
+	if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE &&
 	    !MLX5_CAP_GEN(mdev, cq_period_start_from_cqe))
 		return -EOPNOTSUPP;
 
-	if (!rx_mode_changed)
+	if (!mode_changed)
 		return 0;
 
 	new_channels.params = priv->channels.params;
-	mlx5e_set_rx_cq_mode_params(&new_channels.params, rx_cq_period_mode);
+	if (is_rx_cq)
+		mlx5e_set_rx_cq_mode_params(&new_channels.params, cq_period_mode);
+	else
+		mlx5e_set_tx_cq_mode_params(&new_channels.params, cq_period_mode);
 
 	if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
 		priv->channels.params = new_channels.params;
@@ -1491,6 +1498,16 @@ static int set_pflag_rx_cqe_based_moder(struct net_device *netdev, bool enable)
 	return 0;
 }
 
+static int set_pflag_tx_cqe_based_moder(struct net_device *netdev, bool enable)
+{
+	return set_pflag_cqe_based_moder(netdev, enable, false);
+}
+
+static int set_pflag_rx_cqe_based_moder(struct net_device *netdev, bool enable)
+{
+	return set_pflag_cqe_based_moder(netdev, enable, true);
+}
+
 int mlx5e_modify_rx_cqe_compression_locked(struct mlx5e_priv *priv, bool new_val)
 {
 	bool curr_val = MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_RX_CQE_COMPRESS);
@@ -1578,6 +1595,12 @@ static int mlx5e_set_priv_flags(struct net_device *netdev, u32 pflags)
 	if (err)
 		goto out;
 
+	err = mlx5e_handle_pflag(netdev, pflags,
+				 MLX5E_PFLAG_TX_CQE_BASED_MODER,
+				 set_pflag_tx_cqe_based_moder);
+	if (err)
+		goto out;
+
 	err = mlx5e_handle_pflag(netdev, pflags,
 				 MLX5E_PFLAG_RX_CQE_COMPRESS,
 				 set_pflag_rx_cqe_compress);

drivers/net/ethernet/mellanox/mlx5/core/en_main.c

@@ -681,7 +681,7 @@ static int mlx5e_alloc_rq(struct mlx5e_channel *c,
 	}
 
 	INIT_WORK(&rq->am.work, mlx5e_rx_am_work);
-	rq->am.mode = params->rx_cq_period_mode;
+	rq->am.mode = params->rx_cq_moderation.cq_period_mode;
 	rq->page_cache.head = 0;
 	rq->page_cache.tail = 0;
 
@@ -1974,7 +1974,7 @@ static void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
 	}
 
 	mlx5e_build_common_cq_param(priv, param);
-	param->cq_period_mode = params->rx_cq_period_mode;
+	param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
 }
 
 static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
@@ -1986,8 +1986,7 @@ static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
 	MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
 
 	mlx5e_build_common_cq_param(priv, param);
-
-	param->cq_period_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
+	param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
 }
 
 static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
@@ -3987,14 +3986,32 @@ static bool hw_lro_heuristic(u32 link_speed, u32 pci_bw)
 		 (pci_bw <= 16000) && (pci_bw < link_speed));
 }
 
+void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
+{
+	params->tx_cq_moderation.cq_period_mode = cq_period_mode;
+
+	params->tx_cq_moderation.pkts =
+		MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
+	params->tx_cq_moderation.usec =
+		MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
+
+	if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
+		params->tx_cq_moderation.usec =
+			MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
+
+	MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
+			params->tx_cq_moderation.cq_period_mode ==
+				MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
+}
+
 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
 {
-	params->rx_cq_period_mode = cq_period_mode;
+	params->rx_cq_moderation.cq_period_mode = cq_period_mode;
 
 	params->rx_cq_moderation.pkts =
 		MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
 	params->rx_cq_moderation.usec =
-			MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
+		MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
 
 	if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
 		params->rx_cq_moderation.usec =
@@ -4002,10 +4019,11 @@ void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
 
 	if (params->rx_am_enabled)
 		params->rx_cq_moderation =
-			mlx5e_am_get_def_profile(params->rx_cq_period_mode);
+			mlx5e_am_get_def_profile(cq_period_mode);
 
 	MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
-			params->rx_cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
+			params->rx_cq_moderation.cq_period_mode ==
+				MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
 }
 
 u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
@@ -4065,9 +4083,7 @@ void mlx5e_build_nic_params(struct mlx5_core_dev *mdev,
 			MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
 	params->rx_am_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
 	mlx5e_set_rx_cq_mode_params(params, cq_period_mode);
-
-	params->tx_cq_moderation.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
-	params->tx_cq_moderation.pkts = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
+	mlx5e_set_tx_cq_mode_params(params, cq_period_mode);
 
 	/* TX inline */
 	params->tx_max_inline = mlx5e_get_max_inline_cap(mdev);

drivers/net/ethernet/mellanox/mlx5/core/en_rx_am.c

@@ -63,7 +63,11 @@ profile[MLX5_CQ_PERIOD_NUM_MODES][MLX5E_PARAMS_AM_NUM_PROFILES] = {
 
 static inline struct mlx5e_cq_moder mlx5e_am_get_profile(u8 cq_period_mode, int ix)
 {
-	return profile[cq_period_mode][ix];
+	struct mlx5e_cq_moder cq_moder;
+
+	cq_moder = profile[cq_period_mode][ix];
+	cq_moder.cq_period_mode = cq_period_mode;
+	return cq_moder;
 }
 
 struct mlx5e_cq_moder mlx5e_am_get_def_profile(u8 rx_cq_period_mode)
@@ -75,7 +79,7 @@ struct mlx5e_cq_moder mlx5e_am_get_def_profile(u8 rx_cq_period_mode)
 	else /* MLX5_CQ_PERIOD_MODE_START_FROM_EQE */
 		default_profile_ix = MLX5E_RX_AM_DEF_PROFILE_EQE;
 
-	return profile[rx_cq_period_mode][default_profile_ix];
+	return mlx5e_am_get_profile(rx_cq_period_mode, default_profile_ix);
 }
 
 /* Adaptive moderation logic */