Re: [PATCH net-next v10 2/5] openvswitch: set skb protocol and mac_len when receiving on internal device

On Tue, Jun 07, 2016 at 03:45:27PM -0700, pravin shelar wrote:

On Mon, Jun 6, 2016 at 8:08 PM, Simon Horman [off-list ref] wrote:

quoted

On Thu, Jun 02, 2016 at 03:01:47PM -0700, pravin shelar wrote:

quoted

On Wed, Jun 1, 2016 at 11:24 PM, Simon Horman
[off-list ref] wrote:

quoted

* Set skb protocol based on contents of packet. I have observed this is
  necessary to get actual protocol of a packet when it is injected into an
  internal device e.g. by libnet in which case skb protocol will be set to
  ETH_ALL.

* Set the mac_len which has been observed to not be set up correctly when
  an ARP packet is generated and sent via an openvswitch bridge.
  My test case is a scenario where there are two open vswtich bridges.
  One outputs to a tunnel port which egresses on the other.

The motivation for this is that support for outputting to layer 3 (non-tap)
GRE tunnels as implemented by a subsequent patch depends on protocol and
mac_len being set correctly on receive.

Signed-off-by: Simon Horman <redacted>

---
v10
* Set mac_len

v9
* New patch
---
 net/openvswitch/vport-internal_dev.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/net/openvswitch/vport-internal_dev.c b/net/openvswitch/vport-internal_dev.c
index 2ee48e447b72..f89b1efa88f1 100644
--- a/net/openvswitch/vport-internal_dev.c
+++ b/net/openvswitch/vport-internal_dev.c

@@ -48,6 +48,10 @@ static int internal_dev_xmit(struct sk_buff *skb, struct net_device *netdev)
 {
        int len, err;

+       skb->protocol = eth_type_trans(skb, netdev);
+       skb_push(skb, ETH_HLEN);
+       skb_reset_mac_len(skb);
+

resetting mac-len breaks the assumption about mac_len for referencing
MPLS header ref: skb_mpls_header().

Thanks I had overlooked this. I think it is actually safe as
the mac_len is recalculated quite soon in key_extract() and IIRC
the most important thing is for mac_len to be 0 or non-zero
for the benefit of ovs_flow_key_extract(). None the less it does
seem untidy and moreover inconsistent with the handling in
netdev_port_receive() by a latter patch which does the following:

        eth_type = eth_type_trans(skb, skb->dev);
        skb->mac_len = skb->data - skb_mac_header(skb);
        __skb_push(skb, skb->mac_len);

        if (eth_type == htons(ETH_P_8021Q))
                skb->mac_len += VLAN_HLEN;

Perhaps that logic ought to be in a helper used by both internal_dev_xmit()
and netdev_port_receive(). Or somehow centralised in ovs_vport_receive().

This does looks bit complex. Can we use other skb metadata like
skb_mac_header_was_set()?

Yes, I think that can be made to work if skb->mac_header is unset
for l3 packets in netdev_port_receive(). The following is an incremental
patch on the entire series. Is this the kind of thing you had in mind?

diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
index 86f2cfb19de3..42587d5bf894 100644
--- a/net/openvswitch/flow.c
+++ b/net/openvswitch/flow.c

@@ -729,7 +729,7 @@ int ovs_flow_key_extract(const struct ip_tunnel_info *tun_info,
 	key->phy.skb_mark = skb->mark;
 	ovs_ct_fill_key(skb, key);
 	key->ovs_flow_hash = 0;
-	key->phy.is_layer3 = skb->mac_len == 0;
+	key->phy.is_layer3 = skb_mac_header_was_set(skb) == 0;
 	key->recirc_id = 0;
 
 	err = key_extract(skb, key);

diff --git a/net/openvswitch/vport-internal_dev.c b/net/openvswitch/vport-internal_dev.c
index 484ba529c682..8973d4db509b 100644
--- a/net/openvswitch/vport-internal_dev.c
+++ b/net/openvswitch/vport-internal_dev.c

@@ -50,7 +50,6 @@ static int internal_dev_xmit(struct sk_buff *skb, struct net_device *netdev)
 
 	skb->protocol = eth_type_trans(skb, netdev);
 	skb_push(skb, ETH_HLEN);
-	skb_reset_mac_len(skb);
 
 	len = skb->len;
 	rcu_read_lock();

diff --git a/net/openvswitch/vport-netdev.c b/net/openvswitch/vport-netdev.c
index 3df36df62ee9..4cf3f12ffc99 100644
--- a/net/openvswitch/vport-netdev.c
+++ b/net/openvswitch/vport-netdev.c

@@ -60,22 +60,9 @@ static void netdev_port_receive(struct sk_buff *skb)
 	if (vport->dev->type == ARPHRD_ETHER) {
 		skb_push(skb, ETH_HLEN);
 		skb_postpush_rcsum(skb, skb->data, ETH_HLEN);
-	} else if (vport->dev->type == ARPHRD_NONE) {
-		if (skb->protocol == htons(ETH_P_TEB)) {
-			__be16 eth_type;
-
-			if (unlikely(skb->len < ETH_HLEN))
-				goto error;
-
-			eth_type = eth_type_trans(skb, skb->dev);
-			skb->mac_len = skb->data - skb_mac_header(skb);
-			__skb_push(skb, skb->mac_len);
-
-			if (eth_type == htons(ETH_P_8021Q))
-				skb->mac_len += VLAN_HLEN;
-		} else {
-			skb->mac_len = 0;
-		}
+	} else if (vport->dev->type == ARPHRD_NONE &&
+		   skb->protocol != htons(ETH_P_TEB)) {
+		skb->mac_header = (typeof(skb->mac_header))~0U;
 	}
 
 	ovs_vport_receive(vport, skb, skb_tunnel_info(skb));