Re: [PATCH nf-next 0/8] netfilter: untangle bridge and bridge netfilter

[Florian Westphal <fw@strlen.de>]

Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > nf_bridge is kmalloced blob with some extra information, including
> > the bridge in and outports (mainly for iptables' physdev match).
> > It also has various state bits so we know what manipulations
> > have been performed by bridge netfilter on the skb (e.g.
> > ppp header stripping).
> >
> > nf_bridge also provides scratch space where br_netfilter saves
> > (and later restores) various things, e.g. ipv4 address for
> > dnat detection, mac address to fix up ip fragmented skbs, etc.
> > 
> > But in almost all cases we can avoid using ->data completely.
> 
> I think one of the goals of this patchset is to prepare the removal of
> that nf_bridge pointer from sk_buff which sounds as a good idea to me.
> 
> Did you consider to implement this scratchpad area using a per-cpu
> area? I mean, something similar to what we used to have in
> ip_conntrack for events:

I see that I misread part of what you wrote.

We cannot use percpu data for nf_bridge_info; at least its not trivial
to do (I tried).

Main issues are:
- nfqueue (we have to save/restore info)
- local delivery (skb is enqueued in backlog)
- skb is dropped (we need to reset scratch data)
- DNAT mac hack (skb can be queued in qdisc).

We _can_ use percpu data for passing the original mac
header to the ip_fragment output function.  But, as mentioned,
there is a patch in netdev patchwork that adds extra output
parameter for use with OVS so it seems cleaner to (re-) use common
api in both OVS and br netfilter.

> BTW, 6/8 I think needs some ifdef to make sure NF_CONNTRACK is in
> placed.

Right, thanks for catching this.

> 7/8 needs NF_BRDIGE_MAX_MAC_HEADER_LENGTH which seems to have
> a small typo in it, well these are dependent of 4/8 anyway.

Sorry -- I'm dense -- what typo? :-)