"TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

"TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

[Matthew Hall]

Hello,

I'm just trying to understand what you're supposed to do about this error to 
get the optiomal configuration / performance. The error message and comments 
seem like they're designed for Intel ethernet driver hackers not security 
hackers like myself! ;-)

Note: I'm trying out the Intel 82545EM in VirtualBox because the virtio-net 
approach didn't work, for reasons I enumerated in some previous mails.

    /*
     * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
     * set to 0. If WTHRESH is greater than zero, the RS bit is ignored
     * by the NIC and all descriptors are written back after the NIC
     * accumulates WTHRESH descriptors.
     */
    if (tx_conf->tx_thresh.wthresh != 0 && tx_rs_thresh != 1) {
        RTE_LOG(ERR, PMD, "TX WTHRESH must be set to 0 if "
            "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
            "port=%d queue=%d)\n", (unsigned int)tx_rs_thresh,
                (int)dev->data->port_id, (int)queue_idx);
        return -(EINVAL);
    }

Thanks,
Matthew.

Re: "TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

[Marc Sune]

Which DPDK version are you using

marc

On 19/10/14 00:50, Matthew Hall wrote:
> Hello,
>
> I'm just trying to understand what you're supposed to do about this error to
> get the optiomal configuration / performance. The error message and comments
> seem like they're designed for Intel ethernet driver hackers not security
> hackers like myself! ;-)
>
> Note: I'm trying out the Intel 82545EM in VirtualBox because the virtio-net
> approach didn't work, for reasons I enumerated in some previous mails.
>
>      /*
>       * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
>       * set to 0. If WTHRESH is greater than zero, the RS bit is ignored
>       * by the NIC and all descriptors are written back after the NIC
>       * accumulates WTHRESH descriptors.
>       */
>      if (tx_conf->tx_thresh.wthresh != 0 && tx_rs_thresh != 1) {
>          RTE_LOG(ERR, PMD, "TX WTHRESH must be set to 0 if "
>              "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
>              "port=%d queue=%d)\n", (unsigned int)tx_rs_thresh,
>                  (int)dev->data->port_id, (int)queue_idx);
>          return -(EINVAL);
>      }
>
> Thanks,
> Matthew.

Re: "TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

[Matthew Hall]

1.7.1 with a few minor clang compatibility patches in the example apps
-- 
Sent from my mobile device.

On October 19, 2014 6:46:27 AM PDT, Marc Sune <marc.sune-kpkqNMk1I7M@public.gmane.org> wrote:
>Which DPDK version are you using
>
>marc
>
>On 19/10/14 00:50, Matthew Hall wrote:
>> Hello,
>>
>> I'm just trying to understand what you're supposed to do about this
>error to
>> get the optiomal configuration / performance. The error message and
>comments
>> seem like they're designed for Intel ethernet driver hackers not
>security
>> hackers like myself! ;-)
>>
>> Note: I'm trying out the Intel 82545EM in VirtualBox because the
>virtio-net
>> approach didn't work, for reasons I enumerated in some previous
>mails.
>>
>>      /*
>>       * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
>>       * set to 0. If WTHRESH is greater than zero, the RS bit is
>ignored
>>       * by the NIC and all descriptors are written back after the NIC
>>       * accumulates WTHRESH descriptors.
>>       */
>>      if (tx_conf->tx_thresh.wthresh != 0 && tx_rs_thresh != 1) {
>>          RTE_LOG(ERR, PMD, "TX WTHRESH must be set to 0 if "
>>              "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
>>              "port=%d queue=%d)\n", (unsigned int)tx_rs_thresh,
>>                  (int)dev->data->port_id, (int)queue_idx);
>>          return -(EINVAL);
>>      }
>>
>> Thanks,
>> Matthew.

Re: "TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

[Bruce Richardson]

On Sun, Oct 19, 2014 at 10:08:29AM -0700, Matthew Hall wrote:
> 1.7.1 with a few minor clang compatibility patches in the example apps

Rather than trying to tune the results yourself, maybe you could look to 
merge in the following patch set also, to offload responsibility for the 
driver settings to the driver itself.

http://thread.gmane.org/gmane.comp.networking.dpdk.devel/6189

/Bruce

> -- 
> Sent from my mobile device.
> 
> On October 19, 2014 6:46:27 AM PDT, Marc Sune <marc.sune-kpkqNMk1I7M@public.gmane.org> wrote:
> >Which DPDK version are you using
> >
> >marc
> >
> >On 19/10/14 00:50, Matthew Hall wrote:
> >> Hello,
> >>
> >> I'm just trying to understand what you're supposed to do about this
> >error to
> >> get the optiomal configuration / performance. The error message and
> >comments
> >> seem like they're designed for Intel ethernet driver hackers not
> >security
> >> hackers like myself! ;-)
> >>
> >> Note: I'm trying out the Intel 82545EM in VirtualBox because the
> >virtio-net
> >> approach didn't work, for reasons I enumerated in some previous
> >mails.
> >>
> >>      /*
> >>       * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
> >>       * set to 0. If WTHRESH is greater than zero, the RS bit is
> >ignored
> >>       * by the NIC and all descriptors are written back after the NIC
> >>       * accumulates WTHRESH descriptors.
> >>       */
> >>      if (tx_conf->tx_thresh.wthresh != 0 && tx_rs_thresh != 1) {
> >>          RTE_LOG(ERR, PMD, "TX WTHRESH must be set to 0 if "
> >>              "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
> >>              "port=%d queue=%d)\n", (unsigned int)tx_rs_thresh,
> >>                  (int)dev->data->port_id, (int)queue_idx);
> >>          return -(EINVAL);
> >>      }
> >>
> >> Thanks,
> >> Matthew.
> 

Re: "TX WTHRESH must be set to 0 if tx_rs_thresh is greater than 1"

[Matthew Hall]

On Mon, Oct 20, 2014 at 10:36:01AM +0100, Bruce Richardson wrote:
> On Sun, Oct 19, 2014 at 10:08:29AM -0700, Matthew Hall wrote:
> > 1.7.1 with a few minor clang compatibility patches in the example apps
> 
> Rather than trying to tune the results yourself, maybe you could look to 
> merge in the following patch set also, to offload responsibility for the 
> driver settings to the driver itself.
> 
> http://thread.gmane.org/gmane.comp.networking.dpdk.devel/6189
> 
> /Bruce

Sounds good... I'll try it out and see if I can cherry-pick it onto my 
codebase.

Matthew.