On 11/11/25 3:49 AM, shy liu wrote:
> Hi Dumitru,
> 

Hi Xie Liu,

> Thanks for your review.
> 
> On Mon, Nov 10, 2025 at 7:42 PM Dumitru Ceara <[email protected]> wrote:
>>
>> Hi Xie Liu,
>>
>> Thanks for the patch!
>>
>> On 11/7/25 5:48 AM, [email protected] wrote:
>>> From: Xie Liu <[email protected]>
>>>
>>> Consider the case of stateful Firewall for N-S traffic:
>>>
>>> PUBLIC---S1-(S1-R1)---------(R1-S1)-R1 -------- S2 ---- VM1
>>>
>>> Configuration:
>>>
>>> ovn-nbctl pg-add pg_dgw
>>> ovn-nbctl pg-set-ports pg_dgw S1-R1
>>> ovn-nbctl acl-add pg_dgw from-lport 2000 "inport == @pg_dgw && ip4  && 
>>> icmp4" allow-related
>>> ovn-nbctl acl-add pg_dgw from-lport 1000 "outport == @pg_dgw && ip4" drop
>>> ovn-nbctl acl-add pg_dgw to-lport 1000 "outport == @pg_dgw && ip4" drop
>>> ovn-nbctl lsp-set-options S1-R1 router-port=R1-S1 
>>> enable_router_port_acl=true
>>>
>>> VM1 pings external network.
>>>
>>> Through this patch[1], the ovn-controller assigned a CT zone ID
>>> to the localnet LSP but not the dgw LSP.
>>>
>>> This caused ACL failures: ICMP reply packets from external networks
>>> performed CT lookups in the wrong zone, couldn't match established
>>> connections, and were incorrectly dropped.
>>
>> While I understand why the CT lookup in the wrong zone (in the VIF zone)
>> causes issues, I have my concerns about this behavior change.
>>
>>>
>>> Fix by ensuring ports without CT zone allocation use default zone 0,
>>> preventing incorrect zone inheritance and restoring proper ACL behavior
>>> for distributed gateway scenarios.
>>>
>>
>> Default zone 0 is most likely already used by the host traffic.  Which
>> means we might have unexpected collisions here.  Please see the
>> ovn-kubernetes use case where traffic from the host itself is also sent
>> through conntrack (also in zone 0).
>>
>> Would a correct approach be to instead allocate explicit zone ids for
>> ports that have enable_router_port_acl=true?
> Yes, allocating explicit zone IDs for ports with enable_router_port_acl=true
> is also an option I have considered. It indeed looks more reasonable,
> and I will adopt this approach in the next patch.

Thank you, looking forward to the next revision.

Regards,
Dumitru

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to