Fix checksum related BUG in l2tp_ppp receive path

Hello,

I managed to (reproducably) trigger this bug:

which is -> BUG_ON(offset >= skb_headlen(skb));

(full BUG here: http://www1.aufwind-solutions.de/oopses/pppol2tp1.txt )

Interestingly, the PPP over L2TP (with xl2tpd) is working fine when using my Linux client, but a simple ping from my Android phone causes the BUG on the first packet. It happens on the L2TP -> PPP path, as the ping request never leaves the ppp devices to the local network.

What *might* be different from other setups, is that the machine on which the L2TP packets are decapsulated is a Xen VM, so the packets arrive with CHECKSUM_PARTIAL set.

I followed the code path and saw that in this direction, the checksum members of the SKB are never adjusted, and so my theory is that the packet travels through the ppp stack, back into the IP stack until it tries to leave the network stack to the network device, at which point the checksumming is supposed to happen and the checksum offset is invalid.

I believe that since the L2TP checksum has already been checked and everything, the decapsulated packet is supposedly "fine" and I also saw the following line in l2tp_eth.c:

        /* checksums verified by L2TP */
        skb->ip_summed = CHECKSUM_NONE;

which sounds to mike like the right thing to do, so I modified l2tp_ppp.c accordingly and voila - no more crashes. In case it helps, I would like to share that patch for inclusion or, in case I'm wrong, to ask you to find the correct fix. :)

Fix kernel BUG at net/core/dev.c:2192 (wrong checksum offset) in L2TP -> PPP receive path (skb checksum information is not reset during decapsulation). Reset checksum like it is also done in the l2tp_eth path.

---
 net/l2tp/l2tp_ppp.c | 4 ++++
 1 file changed, 4 insertions(+)

index 637a341..53aabefb 100644
@@ -250,6 +250,10 @@ static void pppol2tp_recv(struct l2tp_session *session, struct sk_buff *skb, int
 		 *   to the inner packet either
 		 */
 		secpath_reset(skb);
+
+		/* checksums verified by L2TP */
+		skb->ip_summed = CHECKSUM_NONE;
+
 		skb_dst_drop(skb);
 		nf_reset(skb);
 

Hi Jana,

Thanks for the patch, and sorry for missing it the first time around!

Could you advise what Android client you're using? I'd like to attempt to reproduce this here. We do a lot of L2TP testing in VM environments (albeit not Xen), so I'm interested as to why we've not hit this.

I think your analysis is correct, but I'm not sure whether setting skb->ip_summed to CHECKSUM_NONE is the right thing to do. It seems like this would disable checksumming for the ICMP echo response. Although that may avoid the BUG, it may not be entirely ideal behaviour.

I'll have a dig into the other tunnelling protocol implementation and see what they do.

Tom Parkin
Katalix Systems Ltd
http://www.katalix.com
Catalysts for your Embedded Linux software development

You've been asked for feedback several days ago, please respond and give people the information they've asked for.

Thanks.