On Wed, Jul 22, 2020 at 11:25 AM Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
>
> On Wed, Jul 22, 2020 at 05:05 PM CEST, Willem de Bruijn wrote:
> > On Wed, Jul 22, 2020 at 11:02 AM Jakub Sitnicki <jakub@xxxxxxxxxxxxxx> wrote:
> >>
> >> On Wed, Jul 22, 2020 at 04:42 PM CEST, Kuniyuki Iwashima wrote:
> >> > Can I submit a patch to net tree that rewrites udp[46]_lib_lookup2() to
> >> > use only 'result' ?
> >>
> >> Feel free. That should make the conflict resolution even easier later
> >> on.
> >
> > Thanks for the detailed analysis, Jakub.
> >
> > Would it be easier to fix this wholly in bpf-next, by introducing
> > reuseport_result there?
>
> Did you mean replicating the Kuniyuki fix in bpf-next, or just
> introducing the intermediate 'reuseport_result' var?
>
> I'm assuming the former, so that the conflict resolving later on will
> reduce to selecting everything from bpf-next side.
Indeed. Since you are already adding a patch to bpf-next to move the
reuseport_has_conns check back. At the same time, it can introduce
reuseport_result:
if (score > badness) {
reuseport_result = lookup_reuseport(net, sk,
skb, saddr, sport, daddr, hnum);
if (reuseport_result && !reuseport_has_conns(sk, false))
return reuseport_result;
result = reuseport_result ? : sk;
badness = score;
}
> TBH, I don't what is the preferred way to handle it. Perhaps DaveM or
> Alexei/Daniel can say what would make their life easiest?
Good point.
With the above, there still remains a merge conflict, of course. But
then we can take bpf-next as is, so I think it would save a separate
patch to net. But not sure whether that helps anything. It does add an
unnecessary variable.
