12/04/2019 17:38, Ferruh Yigit:
> On 4/12/2019 4:07 PM, Thomas Monjalon wrote:
> > 11/03/2019 18:22, Ferruh Yigit:
> >> On 3/4/2019 11:18 AM, David Marchand wrote:
> >>> According to the api, the q_errors[] per queue statistic is for reception
> >>> errors not transmit errors.
> >>> This is a first cleanup on statistics before looking at oerrors.
> >>>
> >>
> >> Yes, the patchset looks aligned with the API documentation [1].
> >>
> >> What can be the solution after cleanup? We can merge this cleanup and
> >> solution
> >> next to each-other to not leave a gap?
> >
> > I think we should merge those fixes in 19.05-rc2.
> >
> > It seems there is a lot more work to achieve on stats, so better
> > to start without waiting for the full picture.
> >
>
> The problem is "q_errors" is available only for Rx queues, and David's patch
> is
> preventing drivers to put Tx error stats into "q_errors" field.
>
> But it is clear that there is a need for a field for Tx queues errors. David
> has
> another patch to using xstats for this. But I believe xstats is making
> solution
> confusing, and now approach is unbalanced for Rx and Tx queues.
>
> I am for adding a new field for Tx queues "q_errors", and this will make
> getting
> stats and David's patch very simple.
>
> The problem with the new fields is it breaks the ABI, but we already increased
> the ABIVER for ethdev this release, I believe this is very good timing for
> this fix.
If changing the stats API, we should increase the number of stats per queue:
#define RTE_ETHDEV_QUEUE_STAT_CNTRS 16
What about 128 queues per port? 256?
