Re: [tcpdump-workers] [PATCH net 1/2] net: dev_queue_xmit_nit: fix skb->vlan_tci field value

[Daniel Borkmann <dborkman@redhat.com>]

On 01/11/2013 03:37 AM, Paul Pearce wrote:
>>> My opinion as a kernel developer is that the network tap is here to have
>>> a copy of the exact frame given to the _device_.

Agreed.

>> Good: as someone who spends lots of time with tcpdump doing both network
>> and protocol diagnostics, it's really important to see exactly there.
>> If that means turning off some hardware offload in order to get the
>> intact 1p header, then that may be fine for many situations.
>> (At 10G, on a live router... well...)
>
> I agree as well.
>
> But I think Ani's point was that for RX packets, as of commit
> bcc6d47903612c3861201cc3a866fb604f26b8b2, the filters are not
> getting exactly what's "on the wire." Independent of hardware
> acceleration the vlan headers are being stripped off and skb->vlan_tci
> is being set. That's was the origin of this whole mess.
>
> The msg from that commit reads in part:
>> Vlan untagging happens early in __netif_receive_skb so the rest of
>> code (ptype_all handlers, rx_handlers) see the skb like it was
>> untagged by hw.
>
> His confusion (which I share) is why it's acceptable to have this
> behavior of removing headers and setting skb->vlan_tci (regardless of
> hardware acceleration) on the RX path but not also set skb->vlan_tci
> on the TX path.
>
> Indepdent of proposed userspace or PACKET_AUXDATA solutions,
> clarification on the RX skb->vlan_tci behavior would be appreciated.
>
> My knowledge of this code is quite limited so it's entirely possible
> I'm off base here. If so please tell me.

While we're at the topic, though it's slightly unrelated this particular
problem, but related to capturing VLANs and ``what's seen on the wire'',
since it was mentioned.

For different NICs/drivers you might get different default behaviours, and
mostly it's the case (I assume, correct me if I'm wrong) that libpcap has
to ``un-untag'' the VLAN headers in user space, doing a memmove(3) for each
stripped VLAN packet in order to ``fix'' this.

Because of this hack, I even got a report of a user recently, that in
Wireshark, he saw a QinQ header, although it should just have been one
VLAN encapsulation (AR8131 driver with ethtool -K eth0 rxvlan off) as he
saw with netsniff-ng (no memmove(3) done there). (I didn't further follow
or verify this report though.)