Skip to contentSkip to conversation

net/mlx5: add non-template flow metadata split

PATCH New Bing Zhao 3 participants 5 messages Jun 13, 2024 – Aug 29, 2024

Subsystems (2)
lore
Share & tools
lore

Bing Zhao

author
#0 lore raw

PATCH 1/1 net/mlx5: add non-template flow metadata split

The method will not create any new flow rule implicitly during split stage, but only generate needed items, actions and attributes based on the detection.

All the rules will still be created in the flow_hw_list_create().

In the meanwhile, once the mark action is specified in the FDB rule, a new rule in the NIC Rx will be created to:
1. match the mark value on REG_C_x in FDB and set it into Rx flow tag field.
2. copy the metadata in REG_C_x' into the REG_B.

If there is no mark, the default rule with only copying metadata will be hit if there is Queue or RSS action in the NIC Rx rule.

Regarding the NIC Tx, only the metadata is relevant and it will be copied in NIC Tx from REG_A into some REG_C_x. The current HWS implementation already has already supported in the default copy rule or the default SQ miss rule in the NIC Tx root table.

---
Depends-on: patch-141122 ("net/mlx5: fix mlx5 device start failure")

---
 drivers/net/mlx5/linux/mlx5_os.c  |  19 +-
 drivers/net/mlx5/meson.build      |   1 +
 drivers/net/mlx5/mlx5.c           |   5 +-
 drivers/net/mlx5/mlx5_flow.c      |  16 +
 drivers/net/mlx5/mlx5_flow.h      |  51 ++-
 drivers/net/mlx5/mlx5_flow_hw.c   | 121 ++++++-
 drivers/net/mlx5/mlx5_nta_split.c | 568 ++++++++++++++++++++++++++++++
 drivers/net/mlx5/mlx5_trigger.c   |   7 +-
 8 files changed, 758 insertions(+), 30 deletions(-)
 create mode 100644 drivers/net/mlx5/mlx5_nta_split.c
drivers/net/mlx5/linux/mlx5_os.c +18 -1
index 50f4810bff..b71f27203c 100644
@@ -503,8 +503,25 @@ mlx5_alloc_shared_dr(struct rte_eth_dev *eth_dev)
 					       mlx5_flex_parser_clone_free_cb);
 	if (!sh->flex_parsers_dv)
 		goto error;
-	if (priv->sh->config.dv_flow_en == 2)
+	if (priv->sh->config.dv_flow_en == 2) {
+		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
+		    sh->dv_regc0_mask) {
+			/* Reuse DV callback functions. */
+			sh->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
+							    MLX5_FLOW_MREG_HTABLE_SZ,
+							    false, true, eth_dev,
+							    flow_nta_mreg_create_cb,
+							    flow_dv_mreg_match_cb,
+							    flow_nta_mreg_remove_cb,
+							    flow_dv_mreg_clone_cb,
+							    flow_dv_mreg_clone_free_cb);
+			if (!sh->mreg_cp_tbl) {
+				err = ENOMEM;
+				goto error;
+			}
+		}
 		return 0;
+	}
 	/* Init port id action list. */
 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
drivers/net/mlx5/meson.build +1 -0
index b279ddf47c..eb5eb2cce7 100644
@@ -52,6 +52,7 @@ if is_linux
             'mlx5_hws_cnt.c',
             'mlx5_flow_quota.c',
             'mlx5_flow_verbs.c',
+            'mlx5_nta_split.c',
     )
 endif
 
drivers/net/mlx5/mlx5.c +2 -3
index e482f7f0e5..654146badf 100644
@@ -2335,8 +2335,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
 	 * If default mreg copy action is removed at the stop stage,
 	 * the search will return none and nothing will be done anymore.
 	 */
-	if (priv->sh->config.dv_flow_en != 2)
-		mlx5_flow_stop_default(dev);
+	mlx5_flow_stop_default(dev);
 	mlx5_traffic_disable(dev);
 	/*
 	 * If all the flows are already flushed in the device stop stage,
@@ -3297,7 +3296,7 @@ mlx5_set_metadata_mask(struct rte_eth_dev *dev)
 		break;
 	case MLX5_XMETA_MODE_META32_HWS:
 		meta = UINT32_MAX;
-		mark = MLX5_FLOW_MARK_MASK;
+		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
 		break;
 	default:
 		meta = 0;
drivers/net/mlx5/mlx5_flow.c +16 -0
index 9ccbbecc50..eb730354e7 100644
@@ -8125,6 +8125,17 @@ mlx5_flow_list_flush(struct rte_eth_dev *dev, enum mlx5_flow_type type,
 void
 mlx5_flow_stop_default(struct rte_eth_dev *dev)
 {
+#ifdef HAVE_MLX5_HWS_SUPPORT
+	struct mlx5_priv *priv = dev->data->dev_private;
+
+	if (priv->sh->config.dv_flow_en == 2) {
+		mlx5_flow_nta_del_default_copy_action(dev);
+		if (!rte_atomic_load_explicit(&priv->hws_mark_refcnt,
+					      rte_memory_order_relaxed))
+			flow_hw_rxq_flag_set(dev, false);
+		return;
+	}
+#endif
 	flow_mreg_del_default_copy_action(dev);
 	flow_rxq_flags_clear(dev);
 }
@@ -8169,7 +8180,12 @@ int
 mlx5_flow_start_default(struct rte_eth_dev *dev)
 {
 	struct rte_flow_error error;
+#ifdef HAVE_MLX5_HWS_SUPPORT
+	struct mlx5_priv *priv = dev->data->dev_private;
 
+	if (priv->sh->config.dv_flow_en == 2)
+		return mlx5_flow_nta_add_default_copy_action(dev, &error);
+#endif
 	/* Make sure default copy action (reg_c[0] -> reg_b) is created. */
 	return flow_mreg_add_default_copy_action(dev, &error);
 }
drivers/net/mlx5/mlx5_flow.h +50 -1

75 lines not on this page · raw message

drivers/net/mlx5/mlx5_flow_hw.c +102 -19

232 lines not on this page · raw message

drivers/net/mlx5/mlx5_nta_split.c new file +568 -0

568 lines not on this page · raw message

drivers/net/mlx5/mlx5_trigger.c +1 -6
index 6fa7c01cd0..1cdf5382c3 100644
@@ -1446,12 +1446,7 @@ mlx5_dev_stop(struct rte_eth_dev *dev)
 	mlx5_mp_os_req_stop_rxtx(dev);
 	rte_delay_us_sleep(1000 * priv->rxqs_n);
 	DRV_LOG(DEBUG, "port %u stopping device", dev->data->port_id);
-	if (priv->sh->config.dv_flow_en == 2) {
-		if (!rte_atomic_load_explicit(&priv->hws_mark_refcnt, rte_memory_order_relaxed))
-			flow_hw_rxq_flag_set(dev, false);
-	} else {
-		mlx5_flow_stop_default(dev);
-	}
+	mlx5_flow_stop_default(dev);
 	/* Control flows for default traffic can be removed firstly. */
 	mlx5_traffic_disable(dev);
 	/* All RX queue flags will be cleared in the flush interface. */
Signature · 1 line
2.34.1
author
#1 lore raw

PATCH v2 1/1 net/mlx5: add metadata split for compatibility API

The method will not create any new flow rule implicitly during split stage, but only generate needed items, actions and attributes based on the detection.

All the rules will still be created in the flow_hw_list_create().

In the meanwhile, once the mark action is specified in the FDB rule, a new rule in the NIC Rx will be created to:
1. match the mark value on REG_C_x in FDB and set it into Rx flow tag field.
2. copy the metadata in REG_C_x' into the REG_B.

If there is no mark, the default rule with only copying metadata will be hit if there is Queue or RSS action in the NIC Rx rule.

Regarding the NIC Tx, only the metadata is relevant and it will be copied in NIC Tx from REG_A into some REG_C_x. The current HWS implementation already has already supported in the default copy rule or the default SQ miss rule in the NIC Tx root table.

---
 drivers/net/mlx5/linux/mlx5_os.c  |  19 +-
 drivers/net/mlx5/meson.build      |   1 +
 drivers/net/mlx5/mlx5.c           |   5 +-
 drivers/net/mlx5/mlx5_flow.c      |  16 +
 drivers/net/mlx5/mlx5_flow.h      |  51 ++-
 drivers/net/mlx5/mlx5_flow_hw.c   | 112 +++++-
 drivers/net/mlx5/mlx5_nta_split.c | 568 ++++++++++++++++++++++++++++++
 drivers/net/mlx5/mlx5_trigger.c   |   7 +-
 8 files changed, 754 insertions(+), 25 deletions(-)
 create mode 100644 drivers/net/mlx5/mlx5_nta_split.c
drivers/net/mlx5/linux/mlx5_os.c +18 -1
index 5e950e9be1..429853e9b2 100644
@@ -503,8 +503,25 @@ mlx5_alloc_shared_dr(struct rte_eth_dev *eth_dev)
 					       mlx5_flex_parser_clone_free_cb);
 	if (!sh->flex_parsers_dv)
 		goto error;
-	if (priv->sh->config.dv_flow_en == 2)
+	if (priv->sh->config.dv_flow_en == 2) {
+		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
+		    sh->dv_regc0_mask) {
+			/* Reuse DV callback functions. */
+			sh->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
+							    MLX5_FLOW_MREG_HTABLE_SZ,
+							    false, true, eth_dev,
+							    flow_nta_mreg_create_cb,
+							    flow_dv_mreg_match_cb,
+							    flow_nta_mreg_remove_cb,
+							    flow_dv_mreg_clone_cb,
+							    flow_dv_mreg_clone_free_cb);
+			if (!sh->mreg_cp_tbl) {
+				err = ENOMEM;
+				goto error;
+			}
+		}
 		return 0;
+	}
 	/* Init port id action list. */
 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
drivers/net/mlx5/meson.build +1 -0
index b279ddf47c..eb5eb2cce7 100644
@@ -52,6 +52,7 @@ if is_linux
             'mlx5_hws_cnt.c',
             'mlx5_flow_quota.c',
             'mlx5_flow_verbs.c',
+            'mlx5_nta_split.c',
     )
 endif
 
drivers/net/mlx5/mlx5.c +2 -3
index 8d266b0e64..cf34766a50 100644
@@ -2335,8 +2335,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
 	 * If default mreg copy action is removed at the stop stage,
 	 * the search will return none and nothing will be done anymore.
 	 */
-	if (priv->sh->config.dv_flow_en != 2)
-		mlx5_flow_stop_default(dev);
+	mlx5_flow_stop_default(dev);
 	mlx5_traffic_disable(dev);
 	/*
 	 * If all the flows are already flushed in the device stop stage,
@@ -3301,7 +3300,7 @@ mlx5_set_metadata_mask(struct rte_eth_dev *dev)
 		break;
 	case MLX5_XMETA_MODE_META32_HWS:
 		meta = UINT32_MAX;
-		mark = MLX5_FLOW_MARK_MASK;
+		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
 		break;
 	default:
 		meta = 0;
drivers/net/mlx5/mlx5_flow.c +16 -0
index 5b3f2b9119..8be1a370cd 100644
@@ -8152,6 +8152,17 @@ mlx5_flow_list_flush(struct rte_eth_dev *dev, enum mlx5_flow_type type,
 void
 mlx5_flow_stop_default(struct rte_eth_dev *dev)
 {
+#ifdef HAVE_MLX5_HWS_SUPPORT
+	struct mlx5_priv *priv = dev->data->dev_private;
+
+	if (priv->sh->config.dv_flow_en == 2) {
+		mlx5_flow_nta_del_default_copy_action(dev);
+		if (!rte_atomic_load_explicit(&priv->hws_mark_refcnt,
+					      rte_memory_order_relaxed))
+			flow_hw_rxq_flag_set(dev, false);
+		return;
+	}
+#endif
 	flow_mreg_del_default_copy_action(dev);
 	flow_rxq_flags_clear(dev);
 }
@@ -8196,7 +8207,12 @@ int
 mlx5_flow_start_default(struct rte_eth_dev *dev)
 {
 	struct rte_flow_error error;
+#ifdef HAVE_MLX5_HWS_SUPPORT
+	struct mlx5_priv *priv = dev->data->dev_private;
 
+	if (priv->sh->config.dv_flow_en == 2)
+		return mlx5_flow_nta_add_default_copy_action(dev, &error);
+#endif
 	/* Make sure default copy action (reg_c[0] -> reg_b) is created. */
 	return flow_mreg_add_default_copy_action(dev, &error);
 }
drivers/net/mlx5/mlx5_flow.h +50 -1

75 lines not on this page · raw message

drivers/net/mlx5/mlx5_flow_hw.c +98 -14

224 lines not on this page · raw message

drivers/net/mlx5/mlx5_nta_split.c new file +568 -0

568 lines not on this page · raw message

drivers/net/mlx5/mlx5_trigger.c +1 -6
index 6fa7c01cd0..1cdf5382c3 100644
@@ -1446,12 +1446,7 @@ mlx5_dev_stop(struct rte_eth_dev *dev)
 	mlx5_mp_os_req_stop_rxtx(dev);
 	rte_delay_us_sleep(1000 * priv->rxqs_n);
 	DRV_LOG(DEBUG, "port %u stopping device", dev->data->port_id);
-	if (priv->sh->config.dv_flow_en == 2) {
-		if (!rte_atomic_load_explicit(&priv->hws_mark_refcnt, rte_memory_order_relaxed))
-			flow_hw_rxq_flag_set(dev, false);
-	} else {
-		mlx5_flow_stop_default(dev);
-	}
+	mlx5_flow_stop_default(dev);
 	/* Control flows for default traffic can be removed firstly. */
 	mlx5_traffic_disable(dev);
 	/* All RX queue flags will be cleared in the flush interface. */
Signature · 1 line
2.34.1
Acked-by Dariusz Sosnowski on v2 1/1
#2 lore raw

Re: [PATCH v2] net/mlx5: add metadata split for compatibility API

Hi,

From: Bing Zhao <[email protected]>
Sent: Thursday, July 4, 2024 1:26 PM
To: Slava Ovsiienko; [email protected]; Raslan Darawsheh
Cc: Ori Kam; Dariusz Sosnowski; Suanming Mou; Matan Azrad
Subject: [PATCH v2] net/mlx5: add metadata split for compatibility API

The method will not create any new flow rule implicitly during split stage, but only generate needed items, actions and attributes based on the detection.

All the rules will still be created in the flow_hw_list_create().

In the meanwhile, once the mark action is specified in the FDB rule, a new rule in the NIC Rx will be created to:
1. match the mark value on REG_C_x in FDB and set it into Rx flow tag field.
2. copy the metadata in REG_C_x' into the REG_B.

If there is no mark, the default rule with only copying metadata will be hit if there is Queue or RSS action in the NIC Rx rule.

Regarding the NIC Tx, only the metadata is relevant and it will be copied in NIC Tx from REG_A into some REG_C_x. The current HWS implementation already has already supported in the default copy rule or the default SQ miss rule in the NIC Tx root table.

Patch applied to next-net-mlx,

Kindest regards,
Raslan Darawsheh

#3 lore raw

Re: [PATCH v2] net/mlx5: add metadata split for compatibility API

09/07/2024 11:36, Raslan Darawsheh:

Quoted 28 lines

Hi,

From: Bing Zhao <[email protected]>
Sent: Thursday, July 4, 2024 1:26 PM
To: Slava Ovsiienko; [email protected]; Raslan Darawsheh
Cc: Ori Kam; Dariusz Sosnowski; Suanming Mou; Matan Azrad
Subject: [PATCH v2] net/mlx5: add metadata split for compatibility API

The method will not create any new flow rule implicitly during split stage, but only generate needed items, actions and attributes based on the detection.

All the rules will still be created in the flow_hw_list_create().

In the meanwhile, once the mark action is specified in the FDB rule, a new rule in the NIC Rx will be created to:
1. match the mark value on REG_C_x in FDB and set it into Rx flow tag field.
2. copy the metadata in REG_C_x' into the REG_B.

If there is no mark, the default rule with only copying metadata will be hit if there is Queue or RSS action in the NIC Rx rule.

Regarding the NIC Tx, only the metadata is relevant and it will be copied in NIC Tx from REG_A into some REG_C_x. The current HWS implementation already has already supported in the default copy rule or the default SQ miss rule in the NIC Tx root table.

Quoted 4 lines

Patch applied to next-net-mlx,

It looks like this patch was missed,
probably not pushed in the branch.

Please send a v3 for integration in 24.11.

Acked-by Dariusz Sosnowski on v2 1/1
#4 lore raw

Re: [PATCH v2] net/mlx5: add metadata split for compatibility API

In reply to Bing Zhao

hi,

From: Bing Zhao <[email protected]>
Sent: Thursday, July 4, 2024 1:26 PM
To: Slava Ovsiienko; [email protected]; Raslan Darawsheh
Cc: Ori Kam; Dariusz Sosnowski; Suanming Mou; Matan Azrad
Subject: [PATCH v2] net/mlx5: add metadata split for compatibility API

The method will not create any new flow rule implicitly during split stage, but only generate needed items, actions and attributes based on the detection.

All the rules will still be created in the flow_hw_list_create().

In the meanwhile, once the mark action is specified in the FDB rule, a new rule in the NIC Rx will be created to:
1. match the mark value on REG_C_x in FDB and set it into Rx flow tag field.
2. copy the metadata in REG_C_x' into the REG_B.

If there is no mark, the default rule with only copying metadata will be hit if there is Queue or RSS action in the NIC Rx rule.

Regarding the NIC Tx, only the metadata is relevant and it will be copied in NIC Tx from REG_A into some REG_C_x. The current HWS implementation already has already supported in the default copy rule or the default SQ miss rule in the NIC Tx root table.

Patch applied to next-net-mlx,

Kindest regards
Raslan Darawsheh