Re: [PATCH] pkt_sched: act_xt support new Xtables interface

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Mon, Dec 24, 2012 at 07:19:43PM +0100, Pablo Neira Ayuso wrote:
> Hi Jamal,
> 
> On Mon, Dec 24, 2012 at 09:05:42AM -0500, Jamal Hadi Salim wrote:
> > On 12-12-24 08:12 AM, Pablo Neira Ayuso wrote:
> > >
> > >conntrack needs to see defragmented packets, you have to call
> > >nf_defrag_ipv4 / _ipv6 respectively before that.
> > >
> > 
> > This should not be too hard to do - although my thinking says this
> > should be a separate action.
> > 
> > >This also changes the semantics of the raw table in iptables since it
> > >will now see packet with conntrack already attached. So this would
> > >also break -j CT --notrack.
> > 
> > Is there a flag we can check which says a flow is not to be tracked?
> > Doesnt nf_conntrack_in() fail if --no track is set?
> 
> The notrack dummy conntrack (consider it a flag) is attached in
> prerouting raw table. By attaching conntracks at ingress, the notrack
> flag will be ignored. Note that this also breaks conntrack templates
> via -j CT, that allows us to set custom conntrack timeouts, zones and
> helpers at prerouting raw.
> 
> Basically, ct templates are attached via -j CT, this template is
> munched by nf_conntrack_in, which adds the corresponding ct features
> based on the template information.

I'm still spinning around this and I don't come with some easy
solution that doesn't break the existing semantics. One possibility
can be to drop the ct reference after leaving ingress, so the lookup
happens again in prerouting after the raw table to attach it again and
no ct is seen in the raw table but:

1) it's suboptimal in case users have rules using ct at ingress and in
   iptables.

2) the conntrack template infrastructure needs to be reworked/replaced
   by something more flexible to attach features to conntracks, so we
   can still attach features for conntrack entries that were created
   at ingress (so helpers / custom timeouts / notrack don't break).