Florian Westphal [off-list ref] wrote:

Eugene Crosser [off-list ref] wrote:

quoted

Maybe a better solution for stray conntrack entries would be to
introduce finer control in netfilter? One possible idea would be to
implement both "track" and "notrack" targets; then a working
configuration would look like this:

'track' is hard to implement correctly because of RELATED traffic.

E.g. 'tcp dport 22 track' won't work correctly because icmp pmtu
won't be handled.

I'd suggest to try a conditional nf_ct_reset that keeps the conntrack
entry if its in another zone.

I can't think of another solution at the moment, the existing behaviour
of resetting conntrack entry for postrouting/output is too old,
otherwise the better solution IMO would be to keep that entry around on
egress if a NAT rewrite has been done. This would avoid the 'double snat'
problem that the 'reset on ingress' tries to solve.

I'm working on this.

Eugene, I think it makes sense if you send a formal revert, a proper
fix for snat+vrf needs more work.

I think this is fixable but it will likely be not acceptable for net
tree.