Today, when an instance sends an ICMPv6 Neighbor Solicitation (NS) packet, it gets sent to ovn-controller which will reply with the appropriate NA.

While this works, it doesn't match the IPv4 behavior for ARP where all the actions are implemented at OVS level without requiring the controller to be up.

This not only seems to create a dependency for the IPv6 dataplane on the controlplane (ovn-controller) but I assume that, depending on the traffic pattern, there could also be a performance hit as all these actions will be processed by a single thread in ovn-controller.

It looks like this has been already requested upstream in the past [0] and there was some limitations but they seem to have been resolved already at userspace [1].
Not sure about kernel level though.

I'll defer to the core OVN team to assess whether this is technically possible at the moment or we need additional BZs/RFEs that target different components.


[0] https://mail.openvswitch.org/pipermail/ovs-discuss/2020-July/050388.html
[1] http://patchwork.ozlabs.org/project/openvswitch/patch/20200928134947.48269-1-fankaixi.li@bytedance.com/