Skip to contentSkip to conversation

macvlan: Only deliver one copy of the frame to the macvlan interface

PATCH In review Alexander Duyck 2 participants 4 messages Oct 8, 2017 – Oct 9, 2017

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 macvlan: Only deliver one copy of the frame to the macvlan interface

From: Alexander Duyck <[email protected]>

This patch intoduces a slight adjustment for macvlan to address the fact that in source mode I was seeing two copies of any packet addressed to the macvlan interface being delivered where there should have been only one.

The issue appears to be that one copy was delivered based on the source MAC address and then the second copy was being delivered based on the destination MAC address. To fix it I am just freeing the second copy instead of delivering it up the stack using the same netdev as was already delivered to.

Fixes 79cf79abce71 ("macvlan: add source mode")
---
 drivers/net/macvlan.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/macvlan.c +2 -1
index d2aea961e0f4..744b0fe6dc78 100644
@@ -484,7 +484,8 @@ static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb)
 		return RX_HANDLER_PASS;
 
 	dev = vlan->dev;
-	if (unlikely(!(dev->flags & IFF_UP))) {
+	if ((vlan->mode == MACVLAN_MODE_SOURCE) ||
+	    unlikely(!(dev->flags & IFF_UP))) {
 		kfree_skb(skb);
 		return RX_HANDLER_CONSUMED;
 	}
#1 lore raw
Quoted 24 lines · Alexander Duyck

From: Alexander Duyck <[email protected]>

This patch intoduces a slight adjustment for macvlan to address the fact that in source mode I was seeing two copies of any packet addressed to the macvlan interface being delivered where there should have been only one.

The issue appears to be that one copy was delivered based on the source MAC address and then the second copy was being delivered based on the destination MAC address. To fix it I am just freeing the second copy instead of delivering it up the stack using the same netdev as was already delivered to.

Fixes 79cf79abce71 ("macvlan: add source mode")
---
 drivers/net/macvlan.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/macvlan.c +0 -0
index d2aea961e0f4..744b0fe6dc78 100644

drivers/net/macvlan.c static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb) · line 491

Quoted 9 lines
drivers/net/macvlan.c +2 -1
@@ -484,7 +484,8 @@ static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb)
 		return RX_HANDLER_PASS;
 
 	dev = vlan->dev;
-	if (unlikely(!(dev->flags & IFF_UP))) {
+	if ((vlan->mode == MACVLAN_MODE_SOURCE) ||
+	    unlikely(!(dev->flags & IFF_UP))) {
 		kfree_skb(skb);
 		return RX_HANDLER_CONSUMED;
 	}

Shouldn't we have a consume_skb() then instead of kfree_skb() ?

We are not really dropping a packet here, only avoiding some artifact cause by the cited commit.

#2 lore raw
Quoted 38 lines · Eric Dumazet
Quoted 24 lines · Alexander Duyck

From: Alexander Duyck <[email protected]>

This patch intoduces a slight adjustment for macvlan to address the fact that in source mode I was seeing two copies of any packet addressed to the macvlan interface being delivered where there should have been only one.

The issue appears to be that one copy was delivered based on the source MAC address and then the second copy was being delivered based on the destination MAC address. To fix it I am just freeing the second copy instead of delivering it up the stack using the same netdev as was already delivered to.

Fixes 79cf79abce71 ("macvlan: add source mode")
---
 drivers/net/macvlan.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/macvlan.c +0 -0
index d2aea961e0f4..744b0fe6dc78 100644

drivers/net/macvlan.c static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb) · line 491

Quoted 9 lines
drivers/net/macvlan.c +2 -1
@@ -484,7 +484,8 @@ static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb)
              return RX_HANDLER_PASS;
 
      dev = vlan->dev;
-     if (unlikely(!(dev->flags & IFF_UP))) {
+     if ((vlan->mode == MACVLAN_MODE_SOURCE) ||
+         unlikely(!(dev->flags & IFF_UP))) {
              kfree_skb(skb);
              return RX_HANDLER_CONSUMED;
      }

Shouldn't we have a consume_skb() then instead of kfree_skb() ?

Quoted 2 lines

We are not really dropping a packet here, only avoiding some artifact cause by the cited commit.

The cited commit basically introduced an issue where we are cloning it and sending the clone to the correct device and then are stuck with the original. The way I fixed it is currently consistent with how broadcast is already being handled for macvlan since they are calling kfree_skb() on the clone that they end up enqueueing for broadcast.

My thought is to look at rewriting this in relation to some other work I am doing, but I wanted to have a fix for net and stable kernels that prevents this frame duplication from occurring. Really in order to handle this correctly my thought is that we should probably be doing a vlan_prev similar to how we have a pt_prev in __netif_receive_skb_core. Then that way when a packet is meant to be handled by one interface, as is the case for most unicast traffic with VLAN regardless of source mode or not we can then just jump back in using RX_HANDLER_ANOTHER.

- Alex

#3 lore raw
Quoted 58 lines · Alexander Duyck
Quoted 38 lines · Eric Dumazet
Quoted 24 lines · Alexander Duyck

From: Alexander Duyck <[email protected]>

This patch intoduces a slight adjustment for macvlan to address the fact that in source mode I was seeing two copies of any packet addressed to the macvlan interface being delivered where there should have been only one.

The issue appears to be that one copy was delivered based on the source MAC address and then the second copy was being delivered based on the destination MAC address. To fix it I am just freeing the second copy instead of delivering it up the stack using the same netdev as was already delivered to.

Fixes 79cf79abce71 ("macvlan: add source mode")
---
 drivers/net/macvlan.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/macvlan.c +0 -0
index d2aea961e0f4..744b0fe6dc78 100644

drivers/net/macvlan.c static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb) · line 491

Quoted 9 lines
drivers/net/macvlan.c +2 -1
@@ -484,7 +484,8 @@ static rx_handler_result_t macvlan_handle_frame(struct sk_buff **pskb)
              return RX_HANDLER_PASS;
 
      dev = vlan->dev;
-     if (unlikely(!(dev->flags & IFF_UP))) {
+     if ((vlan->mode == MACVLAN_MODE_SOURCE) ||
+         unlikely(!(dev->flags & IFF_UP))) {
              kfree_skb(skb);
              return RX_HANDLER_CONSUMED;
      }

Shouldn't we have a consume_skb() then instead of kfree_skb() ?

Quoted 2 lines

We are not really dropping a packet here, only avoiding some artifact cause by the cited commit.

The cited commit basically introduced an issue where we are cloning it and sending the clone to the correct device and then are stuck with the original. The way I fixed it is currently consistent with how broadcast is already being handled for macvlan since they are calling kfree_skb() on the clone that they end up enqueueing for broadcast.

My thought is to look at rewriting this in relation to some other work I am doing, but I wanted to have a fix for net and stable kernels that prevents this frame duplication from occurring. Really in order to handle this correctly my thought is that we should probably be doing a vlan_prev similar to how we have a pt_prev in __netif_receive_skb_core. Then that way when a packet is meant to be handled by one interface, as is the case for most unicast traffic with VLAN regardless of source mode or not we can then just jump back in using RX_HANDLER_ANOTHER.

Quoted 1 line

- Alex

Actually, now that I am thinking it over again maybe us calling kfree_skb() isn't the correct answer. It might make more sense to just return RX_HANDLER_PASS. Then we can defer it to the original interface to drop it.

- Alex