Re: [PATCH net 1/2] packet: do skb_probe_transport_header when we actually have data

[Daniel Borkmann <daniel@iogearbox.net>]

On 11/07/2015 08:06 PM, Eric Dumazet wrote:
> On Sat, 2015-11-07 at 10:53 -0800, Eric Dumazet wrote:
>
>> Well, imagine following scenario (a real one, as I use it all of time,
>> thus how I discovered all trafgen traffic ends up on one slave only)
>>
>> Even if qdisc is bypassed on the bond0, the current handling does not
>> prevent going to the slave qdiscs.
>>
>> So it is not clear to me why we do a selective probe depending on the
>> bypass of first qdisc.
>
> Presumably the transport_header only needs to be set for gso packets in
> some drivers (look at igbvf_tso() for example)
>
> It looks like we might need an audit and/or some guidelines/fixes.

Hmm, yeah, on a (only quick) look, it seems this is mostly needed for the
virtio_net related code in packet_snd() / packet_recvmsg(), not handled in
RX/TX ring paths actually.

$ git grep -n gso_size net/packet/
net/packet/af_packet.c:2748:                    if (vnet_hdr.gso_size == 0)
net/packet/af_packet.c:2825:            skb_shinfo(skb)->gso_size =
net/packet/af_packet.c:2826:                    __virtio16_to_cpu(vio_le(), vnet_hdr.gso_size);
net/packet/af_packet.c:3219:                    vnet_hdr.gso_size =
net/packet/af_packet.c:3220:                            __cpu_to_virtio16(vio_le(), sinfo->gso_size);

Need to take a closer look on Monday.