Wed, Jun 01, 2022 at 12:41:59AM CEST, kuba@kernel.org wrote:

On Tue, 31 May 2022 21:34:42 +0200 Jiri Pirko wrote:

quoted

And again, for the record, I strongly believe that a separate dl
instance for this does not make any sense at all :/ I wonder why you
still think it does.

For purely software reuse reasons. I think the line cards will require
a lot of the same attributes as the full devlink instance, so making
them a subobject which can have all the same attributes is poor SW arch.

Sure, I understand the motivation.

Think about it from OOP perspective, you'd definitely factor all that
stuff out to an abstract class. We can't do that in netlink but whatever
just make it a full dl instance and describe the link between the two.

Most NIC vendors (everyone excluding Netronome?) decided that devlink
instance is equivalent to a bus device which IIUC it was not supposed
to be. It was supposed to be the whole ASIC. If we're okay to stretch

I agree, that is incorrect. That is why I was thinking about sort of
"alias" to make it right (2 PF devlink instances would be one connected
by alias). Not implemented yet though :/

the definition of a dl instance to be "any independently controllable
unit of HW" for NICs then IDK why we can't make a line card a dl
instance.

Well, it is not independently controllable. Well, truth is, that in our
current implementation, there is one independent "configuration", and
that is flash burn of the gearbox. It is done using a "tunnelling"
register which encapsulates register communication what is done during
flash burning.

Are you afraid of hiding dependencies?

Not really, I'm just not sure I see it is worth the excercise.

In czech, we have this saying: "kanon na vrabce". I think that the
following picture is better than any translation :)
https://i.iinfo.cz/images/72/shutterstock-com-kanon-delo-ptak-vrabec-strilet-1.jpg

Will think about it some more.