[dpdk-dev] [pbhagavatula@caviumnetworks.com: Re: [PATCH 2/4] eventdev: add caps API and PMD callbacks for eth Tx adapter]

[dpdk-dev] [pbhagavatula@caviumnetworks.com: Re: [PATCH 2/4] eventdev: add caps API and PMD callbacks for eth Tx adapter]

[Pavan Nikhilesh]

Hi Nikhil,

On Mon, Jul 16, 2018 at 11:25:45AM +0530, Rao, Nikhil wrote:
> On 7/10/2018 4:26 PM, Pavan Nikhilesh wrote:
> > +int __rte_experimental
> > +rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint32_t *caps)
> > +{
> > The caps get API needs to be similar to rx adapter caps get i.e. it needs to
> > have the eth_port_id as a parameter so that the underlying event dev driver can
> > expose INTERNAL PORT capability as not all ethdev drivers have the capability
> > to interact with the eventdevs internal port.
> >
> > rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint16_t eth_port_id,
> >                       uint32_t *caps);
> Hi Pavan,
>
> Is querying the INTERNAL PORT on a per ethdev basis useful to the
> application ?

If an application chooses to use 2 ports one with INTERNAL PORT capability and
one _without_ it then it would be useful to have per ethdev based
classification similar to Rx adapter. The application can classify events based
on event type RTE_EVENT_TYPE_ETHDEV & RTE_EVENT_TYPE_ETH_RX_ADAPTER to
segregate between INTERNAL & NON-INTERNAL port at ingress side and enqueue it
to either rte_event_eth_tx_adapter_enqueue or to the SINGLE link queue
respectively.

Also, I dont think eventdev should iterate over all probed ethdevs and give
the overall caps as an application might want only a specific ethdev to be
connected to the event tx adapter.

>
> For e.g., the txa_init() function in the adapter implementation can only
> decide to use the internal port if it is supported for all ethdevs,
> hence I left that upto the eventdev PMD to decide - i.e., it could
> iterate across txa->dev_count eth devices to make that determination.
> For caps in general, I agree it makes sense to pass in the ethdev, but
> the INTERNAL PORT didn't seem useful on a per ethdev basis.
>
> We could also replace caps_get with something like a
> rte_event_eth_tx_adapter_internal_port_check(dev_id) and add a per
> ethdev caps if needed later.
>
> Thanks,
> Nikhil

Thanks,
Pavan.

Re: [dpdk-dev] [pbhagavatula@caviumnetworks.com: Re: [PATCH 2/4] eventdev: add caps API and PMD callbacks for eth Tx adapter]

[Rao, Nikhil]

> -----Original Message-----
> From: Pavan Nikhilesh [mailto:pbhagavatula@caviumnetworks.com]
> Sent: Monday, July 16, 2018 5:03 PM
> To: Rao, Nikhil <nikhil.rao@intel.com>;
> jkollanukkaran@caviumnetworks.com; olivier.matz@6wind.com
> Cc: dev@dpdk.org
> Subject: [pbhagavatula@caviumnetworks.com: Re: [dpdk-dev] [PATCH 2/4]
> eventdev: add caps API and PMD callbacks for eth Tx adapter]
> 
> Hi Nikhil,
> 
> On Mon, Jul 16, 2018 at 11:25:45AM +0530, Rao, Nikhil wrote:
> > On 7/10/2018 4:26 PM, Pavan Nikhilesh wrote:
> > > +int __rte_experimental
> > > +rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint32_t *caps) {
> > > The caps get API needs to be similar to rx adapter caps get i.e. it
> > > needs to have the eth_port_id as a parameter so that the underlying
> > > event dev driver can expose INTERNAL PORT capability as not all
> > > ethdev drivers have the capability to interact with the eventdevs
> internal port.
> > >
> > > rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint16_t
> eth_port_id,
> > >                       uint32_t *caps);
> > Hi Pavan,
> >
> > Is querying the INTERNAL PORT on a per ethdev basis useful to the
> > application ?
> 
> If an application chooses to use 2 ports one with INTERNAL PORT capability
> and one _without_ it then it would be useful to have per ethdev based
> classification similar to Rx adapter. The application can classify events based
> on event type RTE_EVENT_TYPE_ETHDEV &
> RTE_EVENT_TYPE_ETH_RX_ADAPTER to segregate between INTERNAL &
> NON-INTERNAL port at ingress side and enqueue it to either
> rte_event_eth_tx_adapter_enqueue or to the SINGLE link queue
> respectively.
> 
The current tx adapter is very similar to how the eventdev pipeline app decides between using the generic_tx v/s worker_tx, I guess what you are suggesting is using the 2 concurrently. I am fine with this, 
Would you always assume INTERNAL PORT on the Rx side to deduce INTERNAL PORT on the tx side ? Just curious if that was just an example, in the general case, you could have packets ingressing from an INTERNAL port and egressing on a port that is !INTERNAL ?

> Also, I dont think eventdev should iterate over all probed ethdevs and give
> the overall caps as an application might want only a specific ethdev to be
> connected to the event tx adapter.
> 
Agreed. The current adapter only supports either the generic_tx or the worker_tx models and selects either at the earliest point it is feasible to, 

I will update the patch series for  caps_get()

Thanks,
Nikhil

[dpdk-dev] [PATCH 2/4] eventdev: add caps API and PMD callbacks for eth Tx adapter

[Pavan Nikhilesh]

On Mon, Jul 16, 2018 at 02:17:40PM +0000, Rao, Nikhil wrote:
>
> > -----Original Message-----
> > From: Pavan Nikhilesh [mailto:pbhagavatula@caviumnetworks.com]
> > Sent: Monday, July 16, 2018 5:03 PM
> > To: Rao, Nikhil <nikhil.rao@intel.com>;
> > jkollanukkaran@caviumnetworks.com; olivier.matz@6wind.com
> > Cc: dev@dpdk.org
> > Subject: [pbhagavatula@caviumnetworks.com: Re: [dpdk-dev] [PATCH 2/4]
> > eventdev: add caps API and PMD callbacks for eth Tx adapter]
> >
> > Hi Nikhil,
> >
> > On Mon, Jul 16, 2018 at 11:25:45AM +0530, Rao, Nikhil wrote:
> > > On 7/10/2018 4:26 PM, Pavan Nikhilesh wrote:
> > > > +int __rte_experimental
> > > > +rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint32_t *caps) {
> > > > The caps get API needs to be similar to rx adapter caps get i.e. it
> > > > needs to have the eth_port_id as a parameter so that the underlying
> > > > event dev driver can expose INTERNAL PORT capability as not all
> > > > ethdev drivers have the capability to interact with the eventdevs
> > internal port.
> > > >
> > > > rte_event_eth_tx_adapter_caps_get(uint8_t dev_id, uint16_t
> > eth_port_id,
> > > >                       uint32_t *caps);
> > > Hi Pavan,
> > >
> > > Is querying the INTERNAL PORT on a per ethdev basis useful to the
> > > application ?
> >
> > If an application chooses to use 2 ports one with INTERNAL PORT capability
> > and one _without_ it then it would be useful to have per ethdev based
> > classification similar to Rx adapter. The application can classify events based
> > on event type RTE_EVENT_TYPE_ETHDEV &
> > RTE_EVENT_TYPE_ETH_RX_ADAPTER to segregate between INTERNAL &
> > NON-INTERNAL port at ingress side and enqueue it to either
> > rte_event_eth_tx_adapter_enqueue or to the SINGLE link queue
> > respectively.
> >
> The current tx adapter is very similar to how the eventdev pipeline app decides between using the generic_tx v/s worker_tx, I guess what you are suggesting is using the 2 concurrently. I am fine with this,
> Would you always assume INTERNAL PORT on the Rx side to deduce INTERNAL PORT on the tx side ? Just curious if that was just an example, in the general case, you could have packets ingressing from an INTERNAL port and egressing on a port that is !INTERNAL ?

I was just giving an example :). The application would be free to model the
pipeline accordingly based on event type and caps.

>
> > Also, I dont think eventdev should iterate over all probed ethdevs and give
> > the overall caps as an application might want only a specific ethdev to be
> > connected to the event tx adapter.
> >
> Agreed. The current adapter only supports either the generic_tx or the worker_tx models and selects either at the earliest point it is feasible to,
>
> I will update the patch series for  caps_get()
>
> Thanks,
> Nikhil

Thanks,
Pavan