Re: [PATCH v9 2/8] dt-bindings: Introduce interconnect binding

[Sudeep Holla <sudeep.holla@arm.com>]

On Wed, Oct 03, 2018 at 11:06:45AM -0700, Saravana Kannan wrote:
> 
> 
> On 10/03/2018 02:33 AM, Sudeep Holla wrote:
> > On Tue, Oct 02, 2018 at 11:56:56AM -0700, Saravana Kannan wrote:
> > > On 10/02/2018 04:17 AM, Sudeep Holla wrote:
> > [...]
> > 
> > > > Yes, I do understand I have made the same point multiple time and it's
> > > > intentional. We need to get the fragmented f/w support story fixed.
> > > > Different ARM vendors are doing different things in f/w and ARM sees the
> > > > same fragmentation story as before. We have come up with new specification
> > > > and my annoying multiple emails are just to constantly remind the same.
> > > > 
> > > > I do understand we have existing implementations to consider, but fixing
> > > > the functionality in arbitrary way is not a good design and it better
> > > > to get them fixed for future.
> > > I believe the fragmentation you are referring to is  in the
> > > interface/communication protocol. I see the benefit of standardizing that as
> > > long as the standard actually turns out to be good. But that's completely
> > > separate from what the FW can/can't do. Asking to standardize what the FW
> > > can/can't do doesn't seem realistic as each chip vendor will have different
> > > priorities - power, performance, cost, chip area, etc. It's the conflation
> > > of these separate topics that doesn't help IMHO.
> > I agree on interface/communication protocol fragmentation and firmware
> > can implement whatever the vendor wish. What I was also referring was
> > the mix-n-match approach which should be avoided.
> > 
> > e.g. Device A and B's PM is managed completely by firmware using OSPM hints
> > Suppose Device X's PM is dependent on Device A and B, in which case it's
> > simpler and cleaner to leave Device X PM to firmware. Reading the state
> > of A and B and using that as hint for X is just overhead which firmware
> > can manage better. That was my main concern here: A=CPU and B=some other
> > device and X is inter-connect to which A and B are connected.
> > 
> > If CPU OPPs are obtained from f/w and this inter-connect from DT, mapping
> > then is a mess and that's what I was concerned. I am sorry if that's not
> > the scenario here, I may have mistaken then.
> > 
> What you are asking would be an ideal case, but this is not an ideal world.

Agreed.

> There are tons of constraints for each chip vendor. Saying you can't mix and
> match makes perfect the enemy of the good.

We can have endless debate on that.

> Adding FW support for A and B might make them optimal.

OK...

> But adding support for X might not be possible for
> multiple real world constraints (chip area, cost, time to market, etc).

but is not a good design though. If f/w blindly changes DVFS for X based
on OS request, then there's possibility for clkscrew kind of exploits
still though moving A/B to f/w was to avoid it. The chances are low but
not zero.

> Saying "either do it all or do nothing" is going to hold back a lot progress
> that can come in increments. Heck, we do the same thing in the kernel. We'll
> add basic simple features first and then improve on them. Why is it suddenly
> frowned up if a FW/HW follows the same approach? I'll just have to agree to
> disagree with you on this view point.
>

I agree on adding basic and then improve on that policy. But it's not
fair to compare this mix-'n'-match approach to that. Sorry but I
disagree with the comparison here.

--
Regards,
Sudeep