Re: [PATCH nf-next 1/2] netfilter: nf_tables: add chain to pktinfo structure

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

On Mon, Nov 28, 2016 at 11:32:24AM +0100, Florian Westphal wrote:
> Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > On Mon, Nov 28, 2016 at 01:56:49AM +0100, Florian Westphal wrote:
> > > Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > > > This patch adds the chain object to the pktinfo structure. This
> > > > potentially allow us to know what basechain this packet is walking over
> > > > from the expression evaluation path.
> > > 
> > > ... for what?  Why...?
> > 
> > Quota depletion event notification needs to know from what table
> > delivery is happening, so this one actually belongs to the stateful
> > object patchset..
> 
> Which patch uses this?
> 
> I see nft_chain() call in patch 8, but it doesn't need the chain object
> but uses it to fetch the table pointer.

That's the only client for this new thing so far.

> However, table is available at init() time so this could also be stored
> in ->priv area afaics.
>
> [ I am not opposed to this chain store thing, but after getting rid of
>   a lot of members from pktinfo it seems to me we should not add
>   new ones without a compelling reason ]

OK, nft_pktinfo is still on the 64 bytes cacheline bound. pahole
reports a couple of holes there. Actually we can provide avoid those
holes by reordering. better not to increase pressure there only for
this.

I'll follow a different path: I can store the table pointer in struct
nft_object. Actually, this would be better since I can pass struct
nft_object to obj->type->foo() functions instead of the ugly void * I
have now, then fetch the object data area via something like
nft_data_priv(obj).

Thanks.