Re: [PATCH v4 net-next] MPLS: Use mpls_features to activate software MPLS GSO segmentation

On Mon, Jun 2, 2014 at 5:16 PM, Simon Horman [off-list ref] wrote:

On Mon, Jun 02, 2014 at 05:21:45PM +0100, Thomas Graf wrote:

quoted

On 06/02/14 at 01:43pm, Simon Horman wrote:

quoted

+#ifdef CONFIG_NET_MPLS_GSO
+static netdev_features_t net_mpls_features(struct sk_buff *skb,
+                                      struct net_device *dev,
+                                      netdev_features_t features)
+{
+   /* There is no support for MPLS LRO. So the only way that
+    * an MPLS skb could require GSO segmentation is if it
+    * was received as a non-MPLS skb and then became an MPLS skb.
+    * This may be effected by Open vSwitch in which case the
+    * mac_len will non-zero and not equal to skb_network_offset
+    * as the former indicates the end of L2 while the latter indicates
+    * the beginning of L3 and there is a gap between them occupied
+    * by the MPLS label stack.
+    *
+    * Thus it is possible to avoid traversing any VLAN tags that are
+    * present to determine if the ethtype is MPLS. Instead the
+    * inequality of mac_len and skb_network_offset are used to
+    * determine if a packet is MPLS for the purpose of determining
+    * offload features.
+    */
+   if (skb->mac_len && skb->mac_len != skb_network_offset(skb))
+           features &= dev->mpls_features;
+   return features;
+}

Could you elaborate a bit on the safety of this? What about
GRE GSO which sets mac_len to the inner network offset?

Hi Thomas,

thanks for pointing that out.

It seems to me that I made an error in extending an assumption
that is true inside the (unmerged MPLS patch for) the Open vSwitch
datapath to code outside of the datapath. I had thought this
would be safe as the check should only trigger for packets
manipulated by the datapath.

I now think that its possible that the GRE GSO code could kick in: if the
datapath outputs to GRE. And even if that is not the case it seems to me
that adding an assumption in code in net/core/dev.c to the way mac_len is
set which has not been universally adopted throughout net/ is asking for
trouble.

My _untested_ alternate approach as illustrated below is to check the
ethernet type for MPLS, using skb_network_protocol to account for TEB and
VLANs.

I am slightly concerned about the performance implications of this
approach.  I notice harmonize_features() already makes a call to
skb_network_protocol(). So if performance is a problem perhaps that call
could be leveraged somehow.

To be honest, I think this actually really belongs as part of
netif_skb_features()/harmonize_features(). The point of those
functions is to return the offloading features that are available for
a given packet, so it's not clear why they wouldn't take MPLS into
account. If we merged them then it would both be cleaner and should
avoid any performance issues.