Re: [RFC PATCH v2 net-next 0/3] flow_offload: Re-add per-action statistics

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

On Mon, May 20, 2019 at 04:37:10PM +0100, Edward Cree wrote:
> On 19/05/2019 01:22, Pablo Neira Ayuso wrote:
> > On Fri, May 17, 2019 at 04:27:29PM +0100, Edward Cree wrote:
> >> On 15/05/2019 20:39, Edward Cree wrote:
> > [...]
> >> Pablo, how do the two options interact with your netfilter offload?  I'm
> >>  guessing it's easier for you to find a unique pointer than to generate
> >>  a unique u32 action_index for each action.  I'm also assuming that
> >>  netfilter doesn't have a notion of shared actions.
> > It has that shared actions concept, see:
> >
> > https://netfilter.org/projects/nfacct/
> >
> > Have a look at 'nfacct' in iptables-extensions(8) manpage.
>
> Thanks.  Looking at net/netfilter/nfnetlink_acct.c, it looks as though you
>  don't have a u32 index in there; for the cookie approach, would the
>  address of the struct nf_acct (casted to unsigned long) work to uniquely
>  identify actions that should be shared?
> I'm not 100% sure how nf (or nfacct) offload is going to look, so I might
>  be barking up the wrong tree here.  But it seems like the cookie method
>  should work better for you — even if you did have an index, how would you
>  avoid collisions with TC actions using the same indices if both are in
>  use on a box?  Cookies OTOH are pointers, so guaranteed unique :)

The cookie approach per-action looks fine to me, there's already a
cookie to identify the rule, so this looks natural to me.