Re: [ovs-dev] [PATCH] tests: fix reference output for meter offload stats

[Simon Horman <simon.horman@corigine.com>]

On Wed, Oct 19, 2022 at 02:17:42PM +0200, Ilya Maximets wrote:
> On 10/19/22 10:12, Simon Horman wrote:
> > On Fri, Oct 14, 2022 at 10:40:30AM -0400, Jamal Hadi Salim wrote:
> >> On Fri, Oct 14, 2022 at 9:00 AM Ilya Maximets <i.maximets@ovn.org> wrote:
> >>>
> >>
> >> [..]
> >>>> I thought it was pipe but maybe it is OK(in my opinion that is a bad code
> >>>> for just "count"). We have some (at least NIC) hardware folks on the list.
> >>>
> >>> IIRC, 'OK' action will stop the processing for the packet, so it can
> >>> only be used as a last action in the list.  But we need to count packets
> >>> as a very first action in the list.  So, that doesn't help.
> >>>
> >>
> >> That's why i said it is a bad code - but i believe it's what some of
> >> the hardware
> >> people are doing. Note: it's only bad if you have more actions after because
> >> it aborts the processing pipeline.
> >>
> >>>> Note: we could create an alias to PIPE and call it COUNT if it helps.
> >>>
> >>> Will that help with offloading of that action?  Why the PIPE is not
> >>> offloadable in the first place and will COUNT be offloadable?
> >>
> >> Offloadable is just a semantic choice in this case. If someone is
> >> using OK to count  today - they could should be able to use PIPE
> >> instead (their driver needs to do some transformation of course).
> > 
> > FWIIW, yes, that is my thinking too.
> 
> I don't know that code well, but I thought that tcf_gact_offload_act_setup()
> is a generic function.  And since it explicitly forbids offload of PIPE
> action, no drivers can actually offload it even if they want to.

Sure, but I would expect that can be changed.

> So it's not really a driver's choice in the current kernel code.  Or am I
> missing something?
>
> Best regards, Ilya Maximets.