Re: [RFC PATCH net-next 0/6] devlink: Add device metric support

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 20 Aug 2020 08:35:25 -0600 David Ahern wrote:
> On 8/19/20 12:07 PM, Jakub Kicinski wrote:
> > I don't have a great way forward in mind, sadly. All I can think of is
> > that we should try to create more well defined interfaces and steer
> > away from free-form ones.  
> 
> There is a lot of value in free-form too.

On Tue, 18 Aug 2020 20:35:01 -0700 Jakub Kicinski wrote:
> It's a question of interface, not the value of exposed data.

> > Example, here if the stats are vxlan decap/encap/error - we should
> > expose that from the vxlan module. That way vxlan module defines one
> > set of stats for everyone.
> > 
> > In general unless we attach stats to the object they relate to, we will
> > end up building parallel structures for exposing statistics from the
> > drivers. I posted a set once which was implementing hierarchical stats,
> > but I've abandoned it for this reason.
> > > [...]
> > 
> > IDK. I just don't feel like this is going to fly, see how many names
> > people invented for the CRC error statistic in ethtool -S, even tho
> > there is a standard stat for that! And users are actually parsing the
> > output of ethtool -S to get CRC stats because (a) it became the go-to
> > place for NIC stats and (b) some drivers forget to report in the
> > standard place.
> > 
> > The cover letter says this set replaces the bad debugfs with a good,
> > standard API. It may look good and standard for _vendors_ because they
> > will know where to dump their counters, but it makes very little
> > difference for _users_. If I have to parse names for every vendor I use,
> > I can as well add a per-vendor debugfs path to my script.
> > 
> > The bar for implementation-specific driver stats has to be high.  
> 
> My take away from this is you do not like the names - the strings side
> of it.
> 
> Do you object to the netlink API? The netlink API via devlink?
> 
> 'perf' has json files to describe and document counters
> (tools/perf/pmu-events). Would something like that be acceptable as a
> form of in-tree documentation of counters? (vs Documentation/networking
> or URLs like
> https://community.mellanox.com/s/article/understanding-mlx5-ethtool-counters)

Please refer to what I said twice now about the definition of the stats
exposed here belonging with the VxLAN code, not the driver.

> > Okay, fair. I just think that in datacenter deployments we are way
> > closer to the SDK model than people may want to admit.
> 
> I do not agree with that; the SDK model means you *must* use vendor code
> to make something work. Your argument here is about labels for stats and
> an understanding of their meaning.

Sure, no "must" for passing packets, but you "must" use vendor tooling
to operate a fleet.

Since everybody already has vendor tools what value does this API add?
I still need per vendor logic. Let's try to build APIs which will
actually make user's life easier, which users will want to switch to.