Re: WARNING in __rcu_read_unlock

[Dmitry Vyukov]

On Tue, Dec 18, 2018 at 1:40 PM Stefano Brivio <sbrivio@xxxxxxxxxx> wrote:
>
> On Tue, 18 Dec 2018 09:49:17 +0100
> Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
> > On Tue, Dec 18, 2018 at 12:18 AM Stefano Brivio <sbrivio@xxxxxxxxxx> wrote:
> > >
> > > On Mon, 17 Dec 2018 16:53:36 +0100
> > > Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> > >
> > > > On Mon, Dec 17, 2018 at 4:24 PM Stefano Brivio <sbrivio@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, 17 Dec 2018 06:57:35 -0800
> > > > > Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> > > > >
> > > > > > Might be cause by commit b8a51b38e4d4dec3e379d52c0fe1a66827f7cf1e
> > > > > > fou, fou6: ICMP error handlers for FoU and GUE
> > > > >
> > > > > This:
> > > > >
> > > > > diff --git a/net/ipv4/fou.c b/net/ipv4/fou.c
> > > > > index 0d0ad19ecb87..20a6de26d146 100644
> > > > > --- a/net/ipv4/fou.c
> > > > > +++ b/net/ipv4/fou.c
> > > > > @@ -1008,6 +1008,9 @@ static int gue_err_proto_handler(int proto, struct sk_buff *skb, u32 info)
> > > > >  {
> > > > >         const struct net_protocol *ipprot = rcu_dereference(inet_protos[proto]);
> > > > >
> > > > > +       if (ipprot == IPPROTO_UDP)
> > > > > +               return -EINVAL;
> > > > > +
> > > > >         if (ipprot && ipprot->err_handler) {
> > > > >                 if (!ipprot->err_handler(skb, info))
> > > > >                         return 0;
> > > > >
> > > > > should fix the issue, but I still have to run tests and make sure we
> > > > > don't hit similar cases.
> > > >
> > > > Please don't forget to add a regression test for it too ;)
> > >
> > > Where would you suggest to add this? The only selftest that goes
> >
> > I dunno. But there must be some place for such tests, right?
>
> Not as far as I know. The selftests checking this path, by design, only
> use supported configurations, they don't forge packets.
>
> Maybe it would be nice to have a semi-automated way to isolate and
> describe/name specific conditions found by syzbot via fuzzing and turn
> those into tests that are then repeated periodically. I'm not sure how
> that would look like, but I think it's still more maintainable than a
> pile of C reproducers with forged packets in selftests/net.

It would be nice to do something like this. Filed
https://github.com/google/syzkaller/issues/884
However, there are few open questions that I am not sure how to resolve yet...


> Eric, Cong, Xin, as you also recently fixed a nice deal of similar cases
> reported by syzbot, what do you think? Did you ever feel the need to
> turn a syzbot reproducer into a regression test case?
>
> > > through this path currently is net/pmtu.sh, but as configuration of an
> > > actual UDP-in-GUE tunnel is currently not supported, I would really
> > > need to forge that specific packet, so that doesn't seem to be a good
> > > fit.
> > >
> > > Won't syzbot add this to some list of reproducers that are checked in
> > > the future?
> >
> > It won't. Also fuzzing is complementary to testing, not a replacement:
>
> Indeed, but that doesn't mean we need to limit the potential of fuzzing
> because "it's not testing". It can be used to check for regressions,
> too, especially in these cases.
>
> > https://twitter.com/dvyukov/status/1074719682962358272
>
> Now, I'm extremely thankful for the work you're doing and especially
> for finding this subtle condition with syzbot, but this is quite
> inaccurate. To be exposed to this issue, the user would need to
> have the fou module loaded (it won't autoload), which is used quite
> rarely, and, on top of that, have a UDP tunnel configured. It wouldn't
> have been the kind of "evil packet crashes the internet" scenario you
> were dreaming of ;)

Okay, I see. Full bug assessment is hard. I mess it both ways for
different bugs.
But I did not claim that it does not require some setup :)
And maybe there is somebody important on the internet that uses such
setup. Who knows.