[dpdk-dev] [PATCH] app/testpmd: support unequal number of RXQ and TXQ

[Nipun Gupta]

> -----Original Message-----
> From: Ferruh Yigit <ferruh.yigit at intel.com>
> Sent: Thursday, September 30, 2021 7:16 PM
> To: Nipun Gupta <nipun.gupta at nxp.com>; dev at dpdk.org
> Cc: xiaoyun.li at intel.com; orika at nvidia.com; thomas at monjalon.net; Hemant
> Agrawal <hemant.agrawal at nxp.com>; Jun Yang <jun.yang at nxp.com>
> Subject: Re: [PATCH] app/testpmd: support unequal number of RXQ and TXQ
> 
> On 9/28/2021 12:08 PM, nipun.gupta at nxp.com wrote:
> > From: Jun Yang <jun.yang at nxp.com>
> >
> > The existing forwarding mode configures the total number of
> > queues as the minimum of rxq and txq, so eventually the number
> > of txq are same as rxq.
> > However in some scenarios, specially for flow control the
> > number of rxq and txq can be different.
> > This patch makes the txq and function of rxq for all such
> > scenario instead of keeping 1:1 relationship between the two.
> >
> 
> Hi Nipun,
> 
> I expect more code exists that assumes Rx queue number and Tx queue number
> is
> equal, did you able to test unequal queue numbers with below change?

Hi Ferruh,

Yes, this works with flow, but I will double check on latest code without using
flow too. Also, we will check for any other code which shall need change w.r.t.
this functionality.

> 
> And can you please describe your usecase more, is the device has uneven Rx/Tx
> queues?

Yes, you are correct. In our case, Rx queues which we can use can be much higher
than Tx queues, when rte flow is being used.

> 
> > Signed-off-by: Jun Yang <jun.yang at nxp.com>
> > ---
> >  app/test-pmd/config.c | 6 ++----
> >  1 file changed, 2 insertions(+), 4 deletions(-)
> >
> > diff --git a/app/test-pmd/config.c b/app/test-pmd/config.c
> > index f5765b34f7..7e17f233ba 100644
> > --- a/app/test-pmd/config.c
> > +++ b/app/test-pmd/config.c
> > @@ -3000,8 +3000,6 @@ rss_fwd_config_setup(void)
> >  	int end;
> >
> >  	nb_q = nb_rxq;
> > -	if (nb_q > nb_txq)
> > -		nb_q = nb_txq;
> >  	cur_fwd_config.nb_fwd_lcores = (lcoreid_t) nb_fwd_lcores;
> >  	cur_fwd_config.nb_fwd_ports = nb_fwd_ports;
> >  	cur_fwd_config.nb_fwd_streams =
> > @@ -3038,7 +3036,7 @@ rss_fwd_config_setup(void)
> >  		fs->rx_port = fwd_ports_ids[rxp];
> >  		fs->rx_queue = rxq;
> >  		fs->tx_port = fwd_ports_ids[txp];
> > -		fs->tx_queue = rxq;
> > +		fs->tx_queue = (rxq % nb_txq);
> 
> Is this assumes number of Rx queue is always more than number of Tx queue?

Yes, but on re-thought as per your comment, we shall need to revisit this logic.
I will have a relook on this part too.

Thanks,
Nipun

> 
> >  		fs->peer_addr = fs->tx_port;
> >  		fs->retry_enabled = retry_enabled;
> >  		rxp++;
> > @@ -3253,7 +3251,7 @@ fwd_config_setup(void)
> >  		return;
> >  	}
> >
> > -	if ((nb_rxq > 1) && (nb_txq > 1)){
> > +	if ((nb_rxq > 1) && (nb_txq > 1)) {
> >  		if (dcb_config) {
> >  			for (i = 0; i < nb_fwd_ports; i++) {
> >  				pt_id = fwd_ports_ids[i];
> >