Re: [PATCH net-next] udp_tunnel: properly deal with xfrm gro encap.

[Sabrina Dubroca <sd@queasysnail.net>]

2025-03-19, 16:49:21 +0100, Paolo Abeni wrote:
> 
> 
> On 3/19/25 3:35 PM, Willem de Bruijn wrote:
> > Paolo Abeni wrote:
> >> The blamed commit below does not take in account that xfrm
> >> can enable GRO over UDP encapsulation without going through
> >> setup_udp_tunnel_sock().
> >>
> >> At deletion time such socket will still go through
> >> udp_tunnel_cleanup_gro(), and the failed GRO type lookup will
> >> trigger the reported warning.
> >>
> >> We can safely remove such warning, simply performing no action
> >> on failed GRO type lookup at deletion time.
> >>
> >> Reported-by: syzbot+8c469a2260132cd095c1@syzkaller.appspotmail.com
> >> Closes: https://syzkaller.appspot.com/bug?extid=8c469a2260132cd095c1
> >> Fixes: 311b36574ceac ("udp_tunnel: use static call for GRO hooks when possible")
> >> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> > 
> > Because XFRM does not call udp_tunnel_update_gro_rcv when enabling its
> > UDP GRO offload, from set_xfrm_gro_udp_encap_rcv. But it does call it
> > when disabling the offload, as called for all udp sockest from
> > udp(v6)_destroy_sock. (Just to verify my understanding.)
> 
> Exactly.
> 
> > Not calling udp_tunnel_update_gro_rcv on add will have the unintended
> > side effect of enabling the static call if one other tunnel is also
> > active, breaking UDP GRO for XFRM socket, right?
> 
> Ouch, right again. I think we can/should do better.

We should be able to adapt xfrm to use setup_udp_tunnel_sock, but it's
not a simple conversion because GRO could be enabled separately from
the encap itself. I'm not sure there's much benefit except for a bit
more consistency when we enable the encap with GRO at once (but we'd
still have that odd set_xfrm_gro_udp_encap_rcv to enable GRO after
ESPINUDP has been set up). A few of the UDP encaps that precede
setup_udp_tunnel_sock have been converted, I don't know why ipsec was
left.

-- 
Sabrina