Re: [Intel-wired-lan] bug with rx-udp-gro-forwarding offloading?

[Paolo Abeni <pabeni@redhat.com>]

On Thu, 2023-07-06 at 18:17 +0200, Ian Kumlien wrote:
> On Thu, Jul 6, 2023 at 4:04 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > 
> > On Thu, 2023-07-06 at 15:56 +0200, Eric Dumazet wrote:
> > > On Thu, Jul 6, 2023 at 3:02 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > 
> > > > On Thu, 2023-07-06 at 13:27 +0200, Ian Kumlien wrote:
> > > > > On Thu, Jul 6, 2023 at 10:42 AM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > > > On Wed, 2023-07-05 at 15:58 +0200, Ian Kumlien wrote:
> > > > > > > On Wed, Jul 5, 2023 at 3:29 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > > > > > 
> > > > > > > > On Wed, 2023-07-05 at 13:32 +0200, Ian Kumlien wrote:
> > > > > > > > > On Wed, Jul 5, 2023 at 12:28 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > > > > > > > 
> > > > > > > > > > On Tue, 2023-07-04 at 16:27 +0200, Ian Kumlien wrote:
> > > > > > > > > > > More stacktraces.. =)
> > > > > > > > > > > 
> > > > > > > > > > > cat bug.txt | ./scripts/decode_stacktrace.sh vmlinux
> > > > > > > > > > > [  411.413767] ------------[ cut here ]------------
> > > > > > > > > > > [  411.413792] WARNING: CPU: 9 PID: 942 at include/net/ud     p.h:509
> > > > > > > > > > > udpv6_queue_rcv_skb (./include/net/udp.h:509 net/ipv6/udp.c:800
> > > > > > > > > > > net/ipv6/udp.c:787)
> > > > > > > > > > 
> > > > > > > > > > I'm really running out of ideas here...
> > > > > > > > > > 
> > > > > > > > > > This is:
> > > > > > > > > > 
> > > > > > > > > >         WARN_ON_ONCE(UDP_SKB_CB(skb)->partial_cov);
> > > > > > > > > > 
> > > > > > > > > > sort of hint skb being shared (skb->users > 1) while enqueued in
> > > > > > > > > > multiple places (bridge local input and br forward/flood to tun
> > > > > > > > > > device). I audited the bridge mc flooding code, and I could not find
> > > > > > > > > > how a shared skb could land into the local input path.
> > > > > > > > > > 
> > > > > > > > > > Anyway the other splats reported here and in later emails are
> > > > > > > > > > compatible with shared skbs.
> > > > > > > > > > 
> > > > > > > > > > The above leads to another bunch of questions:
> > > > > > > > > > * can you reproduce the issue after disabling 'rx-gro-list' on the
> > > > > > > > > > ingress device? (while keeping 'rx-udp-gro-forwarding' on).
> > > > > > > > > 
> > > > > > > > > With rx-gro-list off, as in never turned on, everything seems to run fine
> > > > > > > > > 
> > > > > > > > > > * do you have by chance qdiscs on top of the VM tun devices?
> > > > > > > > > 
> > > > > > > > > default qdisc is fq
> > > > > > > > 
> > > > > > > > IIRC libvirt could reset the qdisc to noqueue for the owned tun
> > > > > > > > devices.
> > > > > > > > 
> > > > > > > > Could you please report the output of:
> > > > > > > > 
> > > > > > > > tc -d -s qdisc show dev <tun dev name>
> > > > > > > 
> > > > > > > I don't have these set:
> > > > > > > CONFIG_NET_SCH_INGRESS
> > > > > > > CONFIG_NET_SCHED
> > > > > > > 
> > > > > > > so tc just gives an error...
> > > > > > 
> > > > > > The above is confusing. AS CONFIG_NET_SCH_DEFAULT depends on
> > > > > > CONFIG_NET_SCHED, you should not have a default qdisc, too ;)
> > > > > 
> > > > > Well it's still set in sysctl - dunno if it fails
> > > > > 
> > > > > > Could you please share your kernel config?
> > > > > 
> > > > > Sure...
> > > > > 
> > > > > As a side note, it hasn't crashed - no traces since we did the last change
> > > > 
> > > > It sounds like an encouraging sing! (last famous words...). I'll wait 1
> > > > more day, than I'll submit formally...
> > > > 
> > > > > For reference, this is git diff on the running kernels source tree:
> > > > > diff --git a/net/core/skbuff.c b/net/core/skbuff.c
> > > > > index cea28d30abb5..1b2394ebaf33 100644
> > > > > --- a/net/core/skbuff.c
> > > > > +++ b/net/core/skbuff.c
> > > > > @@ -4270,6 +4270,17 @@ struct sk_buff *skb_segment_list(struct sk_buff *skb,
> > > > > 
> > > > >         skb_push(skb, -skb_network_offset(skb) + offset);
> > > > > 
> > > > > +       if (WARN_ON_ONCE(skb_shared(skb))) {
> > > > > +               skb = skb_share_check(skb, GFP_ATOMIC);
> > > > > +               if (!skb)
> > > > > +                       goto err_linearize;
> > > > > +       }
> > > > > +
> > > > > +       /* later code will clear the gso area in the shared info */
> > > > > +       err = skb_header_unclone(skb, GFP_ATOMIC);
> > > > > +       if (err)
> > > > > +               goto err_linearize;
> > > > > +
> > > > >         skb_shinfo(skb)->frag_list = NULL;
> > > > > 
> > > > >         while (list_skb) {
> > > > 
> > > > ...the above check only, as the other 2 should only catch-up side
> > > > effects of lack of this one. In any case the above address a real
> > > > issue, so we likely want it no-matter-what.
> > > > 
> > > 
> > > Interesting, I wonder if this could also fix some syzbot reports
> > > Willem and I are investigating.
> > > 
> > > Any idea of when the bug was 'added' or 'revealed' ?
> > 
> > The issue specifically addressed above should be present since
> > frag_list introduction commit 3a1296a38d0c ("net: Support GRO/GSO
> > fraglist chaining."). AFAICS triggering it requires non trivial setup -
> > mcast rx on bridge with frag-list enabled and forwarding to multiple
> > ports - so perhaps syzkaller found it later due to improvements on its
> > side ?!?
> 
> I'm also a bit afraid that we just haven't triggered it - i don't see
> any warnings or anything... :/

Let me try to clarify: I hope/think that this chunk alone:

+       /* later code will clear the gso area in the shared info */
+       err = skb_header_unclone(skb, GFP_ATOMIC);
+       if (err)
+               goto err_linearize;
+
        skb_shinfo(skb)->frag_list = NULL;

        while (list_skb) {

does the magic/avoids the skb corruptions -> it everything goes well,
you should not see any warnings at all. Running 'nstat' in the DUT
should give some hints about reaching the relevant code paths.

Cheers,

Paolo