Re: [PATCH v1 net-next] net: dsa: qca: ar9331: add ethtool stats support

[Jakub Kicinski]

On Tue, 17 Nov 2020 01:00:53 +0200 Vladimir Oltean wrote:
> On Mon, Nov 16, 2020 at 02:35:44PM -0800, Jakub Kicinski wrote:
> > On Tue, 17 Nov 2020 00:21:46 +0200 Vladimir Oltean wrote:  
> > > On Mon, Nov 16, 2020 at 01:34:53PM -0800, Jakub Kicinski wrote:  
> > > > You must expose relevant statistics via the normal get_stats64 NDO
> > > > before you start dumping free form stuff in ethtool -S.  
> > >
> > > Completely agree on the point, Jakub, but to be honest we don't give him
> > > that possibility within the DSA framework today, see .ndo_get_stats64 in
> > > net/dsa/slave.c which returns the generic dev_get_tstats64 implementation,
> > > and not something that hooks into the hardware counters, or into the
> > > driver at all, for that matter.  
> >
> > Simple matter of coding, right? I don't see a problem.
> >
> > Also I only mentioned .ndo_get_stats64, but now we also have stats in
> > ethtool->get_pause_stats.  
> 
> Yes, sure we can do that. The pause stats and packet counter ops would
> need to be exposed to the drivers by DSA first, though. Not sure if this
> is something you expect Oleksij to do or if we could pick that up separately
> afterwards.

Well, I feel like unless we draw the line nobody will have 
the incentive to do the work.

I don't mind if it's Oleksij or anyone else doing the plumbing work,
but the task itself seems rather trivial.

> > > But it's good that you raise the point, I was thinking too that we
> > > should do better in terms of keeping the software counters in sync with
> > > the hardware. But what would be a good reference for keeping statistics
> > > on an offloaded interface? Is it ok to just populate the netdev counters
> > > based on the hardware statistics?  
> >
> > IIRC the stats on the interface should be a sum of forwarded in software
> > and in hardware. Which in practice means interface HW stats are okay,
> > given eventually both forwarding types end up in the HW interface
> > (/MAC block).  
> 
> A sum? Wouldn't that count the packets sent/received by the stack twice?

Note that I said _forwarded_. Frames are either forwarded by the HW or
SW (former never hit the CPU, while the latter do hit the CPU or
originate from it).