Re: [PATCH net] net: ensure all external references are released in deferred skbuffs

[Eric Dumazet]

On Wed, Jun 22, 2022 at 4:26 PM Ilya Maximets <i.maximets@xxxxxxx> wrote:
>
> On 6/22/22 13:43, Eric Dumazet wrote:

>
> I tested the patch below and it seems to fix the issue seen
> with OVS testsuite.  Though it's not obvious for me why this
> happens.  Can you explain a bit more?

Anyway, I am not sure we can call nf_reset_ct(skb) that early.

git log seems to say that xfrm check needs to be done before
nf_reset_ct(skb), I have no idea why.

I suspect some incoming packets are not going through
xfrm4_policy_check() and end up being stored in a TCP receive queue.

Maybe something is missing before calling tcp_child_process()


>
> >
> > I note that IPv6 does the nf_reset_ct() earlier, from ip6_protocol_deliver_rcu()
> >
> > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> > index fda811a5251f2d76ac24a036e6b4f4e7d7d96d6f..a06464f96fe0cc94dd78272738ddaab2c19e87db
> > 100644
> > --- a/net/ipv4/tcp_ipv4.c
> > +++ b/net/ipv4/tcp_ipv4.c
> > @@ -1919,6 +1919,8 @@ int tcp_v4_rcv(struct sk_buff *skb)
> >         struct sock *sk;
> >         int ret;
> >
> > +       nf_reset_ct(skb);
> > +
> >         drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
> >         if (skb->pkt_type != PACKET_HOST)
> >                 goto discard_it;
> > @@ -2046,8 +2048,6 @@ int tcp_v4_rcv(struct sk_buff *skb)
> >         if (drop_reason)
> >                 goto discard_and_relse;
> >
> > -       nf_reset_ct(skb);
> > -
> >         if (tcp_filter(sk, skb)) {
> >                 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
> >                 goto discard_and_relse;
>