Merge branch 'gso_forward'

Florian Westphal says:

====================
net: ip: push gso skb forwarding handling down the stack

Turns out doing the segmentation in forwarding was not a bright idea,
there are corner-cases where this has unintended side-effects.

This patch pushes the segmentation downwards.

After this, netif_skb_dev_features() function can be removed
again, it was only added to fetch the features of the output device,
we can just use skb->dev after the pushdown.

Tested with following setup:

host -> kvm_router  -> kvm_host
  mtu 1500        mtu1280

- 'host' has route to kvm_host with locked mtu of 1500
- gso/gro enabled on all interfaces

Did tests with all of following combinations:
- netfilter conntrack off and on on kvm_router
- virtio-net and e1000 driver on kvm_router
- tcp and udp bulk xmit from host to kvm_host

for tcp, I added TCPMSS mangling on kvm_host to make it lie about tcp mss.

Also added a dummy '-t mangle -A POSTROUTING -p udp -f'
rule to make sure no udp fragments are seen in the 'conntrack on'
and 'virtio-net' case.

Also checked (with ping -M do -s 1400)' that it still sends the wanted
icmp error message when size exceeds 1280.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 1 insertions, 6 deletions

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 7ed3a3aa6604..20e99efb1ca6 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3180,12 +3180,7 @@ void netdev_change_features(struct net_device *dev);
 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
 					struct net_device *dev);
 
-netdev_features_t netif_skb_dev_features(struct sk_buff *skb,
-					 const struct net_device *dev);
-static inline netdev_features_t netif_skb_features(struct sk_buff *skb)
-{
-	return netif_skb_dev_features(skb, skb->dev);
-}
+netdev_features_t netif_skb_features(struct sk_buff *skb);
 
 static inline bool net_gso_ok(netdev_features_t features, int gso_type)
 {