On Tue, Nov 23, 2021 at 2:42 PM Eli Cohen <elic at nvidia.com>
wrote:>
> On Tue, Nov 23, 2021 at 10:28:04AM +0800, Jason Wang wrote:
> > On Mon, Nov 22, 2021 at 11:56 PM Parav Pandit <parav at
nvidia.com> wrote:
> > >
> > >
> > >
> > > > From: Eli Cohen <elic at nvidia.com>
> > > > Sent: Monday, November 22, 2021 8:37 PM
> > > >
> > > > On Mon, Nov 22, 2021 at 12:15:36PM +0200, Parav Pandit
wrote:
> > > > >
> > > > >
> > > > > > From: Jason Wang <jasowang at redhat.com>
> > > > > > Sent: Monday, November 22, 2021 3:02 PM
> > > > > >
> > > > > > > If we go with vendor stats, how can we
communicate the information
> > > > > > > to userspace? Currenlty we use netlink
attributes defined to pass
> > > > > > > this information.
> > > > > >
> > > > > > It can be done exactly as what have been done in
the patch, we can
> > > > > > document it as vendor stats.
> > > > > >
> > > > > Yes, attribute to have VENDOR_ prefix in it.
> > > > > >
> > > > > > Ok, I think I get you. So I wonder if it's
more useful to use device
> > > > > > specific counters. For networking, it could be
packets send/received etc.
> > > > >
> > > > > Yes, I equally discussed this previously with Eli as
its more meaningful for end
> > > > users.
> > > > > We just return the device id of it along with queue
number that helps to show
> > > > tx and rx.
> > > > > For ctrl q, it is just ctrl commands and ctrl
completions.
> > > >
> > > > I don't think we should mix send/receive packets for
descriptors statistics. The
> > > > hardware could process a descriptor and still not transmit
any packet.
> > > >
> > > > We can add packets send/recv but descriptor statistics have
their own value.
> > > >
> > > Oh right. I read Jason's comment of _packets_ to fast. I
meant to say send/receive descriptors.
> > > I guess you already named them as tx and rx. Didn't review
the patches in this series yet.
> > >
> > > > To summarize, I can add the VENDOR_ preifx to the attibutes
and re-send or is
> > > > there anything else you think should change?
> > > VENDOR_ prefix and command as iproute2 command as
"vstats" looks fine to me.
> >
> > Ack, but we need to figure out:
> >
> > 1) use descriptors or buffers.
>
> Descriptors.
> Currently we don't support indirect buffers but when we do, we will
> preserve the semantics.
Just to confirm, if I understand correctly:
1) with indirect descriptors, only 1 descriptor is counted
2) with N descriptors chained together, it will report N descriptors
Thanks
>
> > 2) if we use descriptors, for indirect descriptors and descriptor
> > chains how are they counted?
> >
>
> We count descriptors, not buffers.
>
> > Thanks
> >
> > >
> >
>