Re: [dpdk-dev] [PATCH v2 04/11] examples/l3fwd: add ethdev setup based on eventdev

[Jerin Jacob <jerinjacobk@gmail.com>]

On Thu, Jan 2, 2020 at 2:20 PM Nipun Gupta <nipun.gupta@nxp.com> wrote:
>
>
>
> > -----Original Message-----
> > From: Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com>
> > Sent: Thursday, January 2, 2020 11:52 AM
> > To: Nipun Gupta <nipun.gupta@nxp.com>; Jerin Jacob Kollanukkaran
> > <jerinj@marvell.com>; Marko Kovacevic <marko.kovacevic@intel.com>; Ori
> > Kam <orika@mellanox.com>; Bruce Richardson
> > <bruce.richardson@intel.com>; Radu Nicolau <radu.nicolau@intel.com>;
> > Akhil Goyal <akhil.goyal@nxp.com>; Tomasz Kantecki
> > <tomasz.kantecki@intel.com>; Sunil Kumar Kori <skori@marvell.com>;
> > Hemant Agrawal <hemant.agrawal@nxp.com>
> > Cc: dev@dpdk.org
> > Subject: RE: [dpdk-dev] [PATCH v2 04/11] examples/l3fwd: add ethdev setup
> > based on eventdev
> >
> > >> >&local_port_conf);
> > >> >> +               if (ret < 0)
> > >> >> +                       rte_exit(EXIT_FAILURE,
> > >> >> +                                "Cannot configure device: err=%d,
> > >> >> port=%d\n",
> > >> >> +                                ret, port_id);
> > >> >> +
> > >> >
> > >> >We should be using number of RX queues as per the config option
> > >> >provided in the arguments.
> > >> >L3fwd is supposed to support multiple queue. Right?
> > >>
> > >> The entire premise of using event device is to showcase packet
> > >scheduling to
> > >> cores
> > >> without the need for splitting packets across multiple queues.
> > >>
> > >> Queue config is ignored when event mode is selected.
> > >
> > >For atomic queues, we have single queue providing packets to a single
> > >core at a time till processing on that core is completed, irrespective of
> > >the flows on that hardware queue.
> > >And multiple queues are required to distribute separate packets on
> > >separate cores, with these atomic queues maintaining the ordering and
> > >not scheduling on other core, until processing core has completed its
> > >job.
> > >To have this solution generic, we should also take config parameter -
> > >(port, number of queues) to enable multiple ethernet RX queues.
> > >
> >
> > Not sure I follow we connect Rx queue to an event queue which is then
> > linked to multiple event ports which are polled on
> > by respective cores.
>
> This is what we too support, but with atomic queue case the scenario gets little complex.
> Each atomic queue can be scheduled only to one event port at a time, until all the events from
> that event port are processed. Then only it can move to other event port.

This would make it a poll mode. We might as well use normal PMD + RSS
for the same instead.
i.e use l3fwd in poll mode. It will be the same in terms of performance. Right?

>
> To have separate event ports process packets at same time in atomic scenario, multiple queues
> are required. As l3fwd supports multiple queues, it seems legitimate to add the support.
>
> Thanks,
> Nipun
>
> > How would increasing Rx queues help? Distributing flows from single event
> > queue to multiple event ports is the responsibility
> > of Event device as per spec.
> > Does DPAA/2 function differently?
> >
> > Regards,
> > Pavan.
> >
> > >Regards,
> > >Nipun
> > >
> > >>
> > >> >
> > >> >Regards,
> > >> >Nipun
> > >> >
> > >>
> > >> Regards,
> > >> Pavan.