On Sat, Aug 28, 2010 at 08:13:37AM +0800, Herbert Xu wrote:

On Fri, Aug 27, 2010 at 10:50:42PM +0200, Jarek Poplawski wrote:

quoted

After positive netpoll_rx_on() check in vlan_gro_receive() there is
skipped part of the "common" GRO_NORMAL path, especially "pull:" in
dev_gro_receive(), where at least eth header should be copied for
entirely paged skbs. So, eth_type_trans() can read zeroed header only.

Right, thanks for catching this.

quoted

diff --git a/net/8021q/vlan_core.c b/net/8021q/vlan_core.c
index 01ddb04..58289fe 100644
--- a/net/8021q/vlan_core.c
+++ b/net/8021q/vlan_core.c

@@ -139,13 +139,19 @@ gro_result_t vlan_gro_frags(struct napi_struct *napi, struct vlan_group *grp,
 	if (!skb)
 		return GRO_DROP;
 
-	if (netpoll_rx_on(skb)) {
-		skb->protocol = eth_type_trans(skb, skb->dev);
+	/*
+	 * Complete the eth header here, mainly for skb_bond_should_drop(),
+	 * and for netpoll_rx_on() btw.
+	 */
+	skb_gro_pull_in(skb);
+	skb->protocol = eth_type_trans(skb, skb->dev);
+	skb_gro_pull(skb, -ETH_HLEN);

But this code should go into the netpoll (i.e., slow-path) case
only so as not to impede performance.

I'm not sure I can understand. What about skb_bond_should_drop()
mentioned in the comments? IMHO we have to impede performance a bit
(and treat it similarly to non-fragmented path) just to fix it, and
consider optimization of this bonding call later.

Also, we need to fix this for the non-VLAN case as well.

But the only other non-VLAN case I know was fixed already...

Cheers,
Jarek P.