Skip to contentSkip to conversation

net/ixgbe: fix multicast table enable error for VF

PATCH Superseded Wei Zhao 2 participants 5 messages Jan 7, 2019 1 message not in the archive Network drivers

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Wei Zhao

author
#1 lore raw

PATCH v4 1/1 net/ixgbe: fix multicast table enable error for VF

In ixgbe PMD code, all vf ars set with bit IXGBE_VMOLR_ROMPE, which make vf accept packets that match the MTA table,
if some vf update IXGBE_MTA in function ixgbe_vf_set_multicast, then all vf will receive packets from these address.
So there is need to set VMOLR register bit ROPE only after this vf has been set multicast address. If this bit is set when pf host doing initialization, this vf will receive multicast packets with address written in MTA table.And also disable MTA when detect
entry number of MAC address is 0 of configuration from vf. Align to ixgbe pf kernel 5.3.7 code to fix this bug.

Fixes 00e30184daa0 ("ixgbe: add PF support")

---

v2:
change patch name and fix typo in log.

v3:
fix typo and disable MTA when detect
all zero MAC address configuration.

v4:
change code to disable MTA when detect
entry number of MAC address is 0 of configuration from vf.

---
 drivers/net/ixgbe/ixgbe_pf.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/net/ixgbe/ixgbe_pf.c +11 -1
index 4b833ff..4e4d39a 100644
@@ -351,7 +351,7 @@ ixgbe_vf_reset_event(struct rte_eth_dev *dev, uint16_t vf)
 	int rar_entry = hw->mac.num_rar_entries - (vf + 1);
 	uint32_t vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
-	vmolr |= (IXGBE_VMOLR_ROPE | IXGBE_VMOLR_ROMPE |
+	vmolr |= (IXGBE_VMOLR_ROPE |
 			IXGBE_VMOLR_BAM | IXGBE_VMOLR_AUPE);
 	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
 
@@ -503,6 +503,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 	const uint32_t IXGBE_MTA_BIT_MASK = (0x1 << IXGBE_MTA_BIT_SHIFT) - 1;
 	uint32_t reg_val;
 	int i;
+	u32 vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
 	/* Disable multicast promiscuous first */
 	ixgbe_disable_vf_mc_promisc(dev, vf);
@@ -516,6 +517,12 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		vfinfo->vf_mc_hashes[i] = hash_list[i];
 	}
 
+	if (!nb_entries) {
+		vmolr &= ~IXGBE_VMOLR_ROMPE;
+		IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+		return 0;
+	}
+
 	for (i = 0; i < vfinfo->num_vf_mc_hashes; i++) {
 		mta_idx = (vfinfo->vf_mc_hashes[i] >> IXGBE_MTA_BIT_SHIFT)
 				& IXGBE_MTA_INDEX_MASK;
@@ -525,6 +532,9 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		IXGBE_WRITE_REG(hw, IXGBE_MTA(mta_idx), reg_val);
 	}
 
+	vmolr |= IXGBE_VMOLR_ROMPE;
+	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+
 	return 0;
 }
 
Signature · 1 line
2.7.5
author
#2 lore raw

PATCH v5 1/1 net/ixgbe: fix over using multicast table for VF

From: Zhao Wei <[email protected]>

According to the current implementation, all VFs will set bit IXGBE_VMOLR_ROMPE during initialization, this cause any VF will accept packets that match the MTA table. Since the MTA table is shared by all VFs which means if one VF update MTA table in function ixgbe_vf_set_multicast, then all other VFs will receive multicast packets which cause unnecessary performance overhead.

So it's better to set VF's ROPE bit of register VMOLR only if multicast address filter is required on that VF. Also, the ROPE bit should be reset when multicast address filter is requested to clean.

This patch also aligns to the related fix on ixgbe kernel driver 5.3.7.

Fixes 00e30184daa0 ("ixgbe: add PF support")

---

v2:
change patch name and fix typo in log.

v3:
fix typo and disable MTA when detect
all zero MAC address configuration.

v4:
change code to disable MTA when detect
entry number of MAC address is 0 of configuration from vf.

v5:
change git log comment, title and code format.

---
 drivers/net/ixgbe/ixgbe_pf.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/net/ixgbe/ixgbe_pf.c +11 -1
index 4b833ff..be0c076 100644
@@ -351,7 +351,7 @@ ixgbe_vf_reset_event(struct rte_eth_dev *dev, uint16_t vf)
 	int rar_entry = hw->mac.num_rar_entries - (vf + 1);
 	uint32_t vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
-	vmolr |= (IXGBE_VMOLR_ROPE | IXGBE_VMOLR_ROMPE |
+	vmolr |= (IXGBE_VMOLR_ROPE |
 			IXGBE_VMOLR_BAM | IXGBE_VMOLR_AUPE);
 	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
 
@@ -503,6 +503,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 	const uint32_t IXGBE_MTA_BIT_MASK = (0x1 << IXGBE_MTA_BIT_SHIFT) - 1;
 	uint32_t reg_val;
 	int i;
+	u32 vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
 	/* Disable multicast promiscuous first */
 	ixgbe_disable_vf_mc_promisc(dev, vf);
@@ -516,6 +517,12 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		vfinfo->vf_mc_hashes[i] = hash_list[i];
 	}
 
+	if (nb_entries == 0) {
+		vmolr &= ~IXGBE_VMOLR_ROMPE;
+		IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+		return 0;
+	}
+
 	for (i = 0; i < vfinfo->num_vf_mc_hashes; i++) {
 		mta_idx = (vfinfo->vf_mc_hashes[i] >> IXGBE_MTA_BIT_SHIFT)
 				& IXGBE_MTA_INDEX_MASK;
@@ -525,6 +532,9 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		IXGBE_WRITE_REG(hw, IXGBE_MTA(mta_idx), reg_val);
 	}
 
+	vmolr |= IXGBE_VMOLR_ROMPE;
+	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+
 	return 0;
 }
 
Signature · 1 line
2.7.5
author
#3 lore raw

PATCH v5 1/1 net/ixgbe: fix over using multicast table for VF

In reply to Wei Zhao

According to the current implementation, all VFs will set bit IXGBE_VMOLR_ROMPE during initialization, this cause any VF will accept packets that match the MTA table. Since the MTA table is shared by all VFs which means if one VF update MTA table in function ixgbe_vf_set_multicast, then all other VFs will receive multicast packets which cause unnecessary performance overhead.

So it's better to set VF's ROPE bit of register VMOLR only if multicast address filter is required on that VF. Also, the ROPE bit should be reset when multicast address filter is requested to clean.

This patch also aligns to the related fix on ixgbe kernel driver 5.3.7.

Fixes 00e30184daa0 ("ixgbe: add PF support")

---

v2:
change patch name and fix typo in log.

v3:
fix typo and disable MTA when detect
all zero MAC address configuration.

v4:
change code to disable MTA when detect
entry number of MAC address is 0 of configuration from vf.

v5:
change git log comment, title and code format.

---
 drivers/net/ixgbe/ixgbe_pf.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/net/ixgbe/ixgbe_pf.c +11 -1
index 4b833ff..be0c076 100644
@@ -351,7 +351,7 @@ ixgbe_vf_reset_event(struct rte_eth_dev *dev, uint16_t vf)
 	int rar_entry = hw->mac.num_rar_entries - (vf + 1);
 	uint32_t vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
-	vmolr |= (IXGBE_VMOLR_ROPE | IXGBE_VMOLR_ROMPE |
+	vmolr |= (IXGBE_VMOLR_ROPE |
 			IXGBE_VMOLR_BAM | IXGBE_VMOLR_AUPE);
 	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
 
@@ -503,6 +503,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 	const uint32_t IXGBE_MTA_BIT_MASK = (0x1 << IXGBE_MTA_BIT_SHIFT) - 1;
 	uint32_t reg_val;
 	int i;
+	u32 vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
 	/* Disable multicast promiscuous first */
 	ixgbe_disable_vf_mc_promisc(dev, vf);
@@ -516,6 +517,12 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		vfinfo->vf_mc_hashes[i] = hash_list[i];
 	}
 
+	if (nb_entries == 0) {
+		vmolr &= ~IXGBE_VMOLR_ROMPE;
+		IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+		return 0;
+	}
+
 	for (i = 0; i < vfinfo->num_vf_mc_hashes; i++) {
 		mta_idx = (vfinfo->vf_mc_hashes[i] >> IXGBE_MTA_BIT_SHIFT)
 				& IXGBE_MTA_INDEX_MASK;
@@ -525,6 +532,9 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev, uint32_t vf, uint32_t *msgbuf)
 		IXGBE_WRITE_REG(hw, IXGBE_MTA(mta_idx), reg_val);
 	}
 
+	vmolr |= IXGBE_VMOLR_ROMPE;
+	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+
 	return 0;
 }
 
Signature · 1 line
2.7.5
Acked-by on v5 1/1 Applied
#4 lore raw

Re: [PATCH v5] net/ixgbe: fix over using multicast table for VF

Quoted 21 lines

-----Original Message-----
From: Zhao1, Wei
Sent: Monday, January 7, 2019 3:23 PM
To: [email protected]
Cc: [email protected]; Zhang, Qi Z <[email protected]>; Wu, Jingjing <[email protected]>; Zhao1, Wei <[email protected]>
Subject: [PATCH v5] net/ixgbe: fix over using multicast table for VF

According to the current implementation, all VFs will set bit IXGBE_VMOLR_ROMPE during initialization, this cause any VF will accept packets that match the MTA table. Since the MTA table is shared by all VFs which means if one VF update MTA table in function ixgbe_vf_set_multicast, then all other VFs will receive multicast packets which cause unnecessary performance overhead.

So it's better to set VF's ROPE bit of register VMOLR only if multicast address filter is required on that VF.
Also, the ROPE bit should be reset when multicast address filter is requested to clean.

Quoted 5 lines

This patch also aligns to the related fix on ixgbe kernel driver 5.3.7.

Fixes 00e30184daa0 ("ixgbe: add PF support")

Applied to dpdk-next-net-intel.

Thanks
Qi

#5 lore raw
In reply to Wei Zhao
Quoted 9 lines

-----Original Message-----
From: Zhao1, Wei
Sent: Monday, January 7, 2019 2:16 PM
To: [email protected]
Cc: [email protected]; Zhang, Qi Z <[email protected]>; Wu, Jingjing <[email protected]>; Zhao1, Wei <[email protected]>
Subject: [PATCH v4] net/ixgbe: fix multicast table enable error for VF

How about change the title to:

"fix over using multicast table for VF"

?

Quoted 10 lines

In ixgbe PMD code, all vf ars set with bit IXGBE_VMOLR_ROMPE, which make vf accept packets that match the MTA table, if some vf update IXGBE_MTA in function ixgbe_vf_set_multicast, then all vf will receive packets from these address.
So there is need to set VMOLR register bit ROPE only after this vf has been set multicast address. If this bit is set when pf host doing initialization, this vf will receive multicast packets with address written in MTA table.And also disable MTA when detect entry number of MAC address is 0 of configuration from vf. Align to ixgbe pf kernel 5.3.7 code to fix this bug.

I did below re-word to make it easy to understand.

According to the current implementation, all VFs will set bit IXGBE_VMOLR_ROMPE during initialization, this cause any VF will accept packets that match the MTA table. Since the MTA table is shared by all VFs which means if one VF update MTA table in function ixgbe_vf_set_multicast, then all other VFs will receive multicast packets which cause unnecessary performance overhead.

So it's better to set VF's ROPE bit of register VMOLR only if multicast address filter is required on that VF. Also, the ROPE bit should be reset when multicast address filter is requested to clean.

This patch also aligns to the related fix on ixgbe kernel driver 5.3.7.

Quoted 46 lines
Fixes 00e30184daa0 ("ixgbe: add PF support")

---

v2:
change patch name and fix typo in log.

v3:
fix typo and disable MTA when detect
all zero MAC address configuration.

v4:
change code to disable MTA when detect
entry number of MAC address is 0 of configuration from vf.

---
 drivers/net/ixgbe/ixgbe_pf.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/net/ixgbe/ixgbe_pf.c index renamed from drivers/net/ixgbe/ixgbe_pf.c +0 -0

4b833ff..4e4d39a 100644

drivers/net/ixgbe/ixgbe_pf.c +2 -1
@@ -351,7 +351,7 @@ ixgbe_vf_reset_event(struct rte_eth_dev *dev, uint16_t
 vf)
 	int rar_entry = hw->mac.num_rar_entries - (vf + 1);
 	uint32_t vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
-	vmolr |= (IXGBE_VMOLR_ROPE | IXGBE_VMOLR_ROMPE |
+	vmolr |= (IXGBE_VMOLR_ROPE |
 			IXGBE_VMOLR_BAM | IXGBE_VMOLR_AUPE);
 	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
@@ -503,6 +503,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev,
 uint32_t vf, uint32_t *msgbuf)
 	const uint32_t IXGBE_MTA_BIT_MASK = (0x1 << IXGBE_MTA_BIT_SHIFT) -
 1;
 	uint32_t reg_val;
 	int i;
+	u32 vmolr = IXGBE_READ_REG(hw, IXGBE_VMOLR(vf));
 
drivers/net/ixgbe/ixgbe_pf.c +0 -0
 	/* Disable multicast promiscuous first */
 	ixgbe_disable_vf_mc_promisc(dev, vf);

drivers/net/ixgbe/ixgbe_pf.c ixgbe_vf_set_multicast(struct rte_eth_dev *dev, · line 521

Quoted 5 lines
drivers/net/ixgbe/ixgbe_pf.c +1 -0
@@ -516,6 +517,12 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev,
 uint32_t vf, uint32_t *msgbuf)
 		vfinfo->vf_mc_hashes[i] = hash_list[i];
 	}
 
+	if (!nb_entries) {

It's better to write as

	If (nb_entries == 0)

to follow the coding guideline.

Quoted 21 lines
drivers/net/ixgbe/ixgbe_pf.c +5 -0
+		vmolr &= ~IXGBE_VMOLR_ROMPE;
+		IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+		return 0;
+	}
+
 	for (i = 0; i < vfinfo->num_vf_mc_hashes; i++) {
 		mta_idx = (vfinfo->vf_mc_hashes[i] >> IXGBE_MTA_BIT_SHIFT)
drivers/net/ixgbe/ixgbe_pf.c +3 -1
 				& IXGBE_MTA_INDEX_MASK;
@@ -525,6 +532,9 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev,
 uint32_t vf, uint32_t *msgbuf)
 		IXGBE_WRITE_REG(hw, IXGBE_MTA(mta_idx), reg_val);
 	}
 
+	vmolr |= IXGBE_VMOLR_ROMPE;
+	IXGBE_WRITE_REG(hw, IXGBE_VMOLR(vf), vmolr);
+
 	return 0;
 }
 
--

2.7.5