On Fri, May 20, 2016, at 03:56, Cong Wang wrote:
> On Thu, May 19, 2016 at 3:50 PM, Hannes Frederic Sowa
> <han...@stressinduktion.org> wrote:
> > Hi Cong,
> >
> > On Fri, May 20, 2016, at 00:33, Cong Wang wrote:
> >> On Thu, May 19, 2016 at 6:58 AM, Hannes Frederic Sowa
> >> <han...@stressinduktion.org> wrote:
> >> > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> >> > index 2e3ebfe5549ef5..d56c0559b477cb 100644
> >> > --- a/net/ipv4/udp.c
> >> > +++ b/net/ipv4/udp.c
> >> > @@ -1565,7 +1565,7 @@ int udp_queue_rcv_skb(struct sock *sk, struct
> >> > sk_buff *skb)
> >> >
> >> > /* if we're overly short, let UDP handle it */
> >> > encap_rcv = ACCESS_ONCE(up->encap_rcv);
> >> > - if (skb->len > sizeof(struct udphdr) && encap_rcv) {
> >> > + if (encap_rcv) {
> >>
> >>
> >> I don't think you can just remove it here, l2tp_udp_recv_core() still
> >> relies on it:
> >>
> >> /* UDP has verifed checksum */
> >>
> >> /* UDP always verifies the packet length. */
> >> __skb_pull(skb, sizeof(struct udphdr));
> >
> > I think this is fine, we check on every entrance to udp that we may pull
> > (pskb_may_pull) an udphdr but we really never pull the header. At this
> > point we are guaranteed to have skb->len of at least sizeof(struct
> > udphdr).
>
> Ah, yeah, __udp4_lib_rcv().
Yep, or the respective IPv6 part.
> BTW, this part (the second half) of your patch is an optimization,
> the rest (the first half) is a real bug fix. It makes sense to split
> them if you plan to backport it to -stable.
The problem in the second part is that it is a '>' comparison and not a
'>=' check. Thus if you hit the socket with a UDP packet with no payload
after the UDP header it also gets enqueued in the socket queue of vxlan
or geneve.
Bye,
Hannes
