Re: [PATCH net] bareudp: Fix device stats updates.

[Guillaume Nault <gnault@redhat.com>]

On Wed, Sep 04, 2024 at 07:50:55PM -0600, David Ahern wrote:
> On 9/4/24 11:54 AM, Guillaume Nault wrote:
> > [Adding David Ahern for the vrf/dstats discussion]
> > 
> > On Wed, Sep 04, 2024 at 07:57:32AM -0700, Jakub Kicinski wrote:
> >> On Wed, 4 Sep 2024 14:29:44 +0200 Guillaume Nault wrote:
> >>>> The driver already uses struct pcpu_sw_netstats, would it make sense to
> >>>> bump it up to struct pcpu_dstats and have per CPU rx drops as well?
> >>>
> >>> Long term, I was considering moving bareudp to use dev->tstats for
> >>> packets/bytes and dev->core_stats for drops. It looks like dev->dstats
> >>> is only used for VRF, so I didn't consider it.
> >>
> >> Right, d stands for dummy so I guess they also were used by dummy
> >> at some stage? Mostly I think it's a matter of the other stats being
> >> less recent.
> > 
> > Looks like dummy had its own dstats, yes. But those dstats were really
> > like the current lstats (packets and bytes counters, nothing for
> > drops). Dummy was later converted to lstats by commit 4a43b1f96b1d
> > ("net: dummy: use standard dev_lstats_add() and dev_lstats_read()").
> > 
> > The dstats we have now really come from vrf (different counters for tx
> > and rx and counters for packet drops), which had its own implementation
> > at that time.
> > 
> > My understanding is that vrf implemented its own dstats in order to
> > have per-cpu counters for regular bytes/packets counters and also for
> > packet drops.
> 
> VRF was following other per-cpu counters that existed in 2015-2016
> timeframe.

Thanks. That was my impression as well.

> I have no preference on the naming; just wanted per-cpu counters.
>