On Thu, Sep 30, 2021 at 08:08:53AM +0200, Daniel Borkmann wrote:
> On 9/28/21 11:55 AM, Pablo Neira Ayuso wrote:
> > Hi,
> >
> > This patchset v5 that re-adds the Netfilter egress:
> >
> > 1) Rename linux/netfilter_ingress.h to linux/netfilter_netdev.h
> > from Lukas Wunner.
> >
> > 2) Generalize ingress hook file to accomodate egress support,
> > from Lukas Wunner.
> >
> > 3) Modularize Netfilter ingress hook into nf_tables_netdev: Daniel
> > Borkmann is requesting for a mechanism to allow to blacklist
> > Netfilter, this allows users to blacklist this new module that
> > includes ingress chain and the new egress chain for the netdev
> > family. There is no other in-tree user of the ingress and egress
> > hooks than this which might interfer with his matter.
> >
> > 4) Place the egress hook again before the tc egress hook as requested
> > by Daniel Borkmann. Patch to add egress hook from Lukas Wunner.
> > The Netfilter egress hook remains behind the static key, if unused
> > performance degradation is negligible.
> >
> > 5) Add netfilter egress handling to af_packet.
> >
> > Arguably, distributors might decide to compile nf_tables_netdev
> > built-in. Traditionally, distributors have compiled their kernels using
> > the default configuration that Netfilter Kconfig provides (ie. use
> > modules whenever possible). In any case, I consider that distributor
> > policy is out of scope in this discussion, providing a mechanism to
> > allow Daniel to prevent Netfilter ingress and egress chains to be loaded
> > should be sufficient IMHO.
>
> Hm, so in the case of SRv6 users were running into a similar issue and commit
> 7a3f5b0de364 ("netfilter: add netfilter hooks to SRv6 data plane") [0] added
> a new hook along with a sysctl which defaults the new hook to off.
>
> The rationale for it was given as "the hooks are enabled via nf_hooks_lwtunnel
> sysctl to make sure existing netfilter rulesets do not break." [0,1]
>
> If the suggestion to flag the skb [2] one way or another from the tc forwarding
> path (e.g. skb bit or per-cpu marker) is not technically feasible, then why not
> do a sysctl toggle like in the SRv6 case?
I am already providing a global toggle to disable netdev
ingress/egress hooks?
In the SRv6 case that is not possible.
Why do you need you need a sysctl knob when my proposal is already
addressing your needs?
