Re: Problems with fragments since gso skb forwarding changes in virtual environment

[Florian Westphal <fw@strlen.de>]

Florian Westphal <fw@strlen.de> wrote:

[ cc'd Eric ]

> Thanks for clarifying. In this case there is another problem as well as
> no fragments should be generated in the forwarding path if the outgoing mtu
> is not reduced.
> 
> Most likely a problem with udp gso + skb_gso_network_seglen().

Indeed, the problem is here:

unsigned int skb_gso_transport_seglen(const struct sk_buff *skb)
{
       const struct skb_shared_info *shinfo = skb_shinfo(skb);
       unsigned int hdr_len;

       if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
               hdr_len = tcp_hdrlen(skb);
       else
               hdr_len = sizeof(struct udphdr);
       return hdr_len + shinfo->gso_size;
}

For large GSO UDP packet received via virtio shinfo->gso_size is 1480 (mtu 1500).

Thus skb_gso_transport_seglen() claims transport seglen of 1488 which makes fwd
path think we're trying to send 1508 byte segments.

I am not familiar with UFO; setting hdr_len = 0 in else clause makes things
work for me.  I'll see if the sizeof(struct udphdr) is bogus for ufo or if ->gso_size
is wrong.