[PATCH v3 1/3] dt-bindings: phy: qcom,qmp: Add PCIe qcom,refclk-always-on property

[Manivannan Sadhasivam]

On Thu, Dec 07, 2023 at 02:36:49PM +0100, Johan Hovold wrote:
> On Thu, Dec 07, 2023 at 06:50:32PM +0530, Manivannan Sadhasivam wrote:
> > On Thu, Dec 07, 2023 at 02:05:12PM +0100, Johan Hovold wrote:
> > > On Thu, Dec 07, 2023 at 03:42:52PM +0530, Manivannan Sadhasivam wrote:
> > > > On Thu, Dec 07, 2023 at 10:51:09AM +0100, Johan Hovold wrote:
> 
> > > > > Shouldn't that be
> > > > > 
> > > > > 	qcom,broken-clkreq
> > > > > 
> > > > > since its the CLKREQ# signal used to request REFCLK that is broken, not
> > > > > the REFCLK itself?
> > > > > 
> > > > 
> > > > Darn... You are right. I got carried away by the initial property name. Thanks
> > > > for spotting!
> > > 
> > > Thinking some more on this after hitting send: It may still be wrong
> > > with a 'broken-clkreq' property in the PHY instead of in the controller
> > > (or endpoint).
> > > 
> > > Could there not be other ways to handle a broken clkreq signal so that
> > > this really should be a decision made by the OS, for example, to disable
> > > L1 substates and clock PM?
> > 
> > One has to weigh the power consumption between keeping refclk always on and
> > disabling L1SS. Chaitanya, can you measure power consumption in both cases?
> 
> Sure, my point was just that that's a policy decision and not something
> that should be encoded in the devicetree (as was initially proposed).
> 
> And that the right place for the renamed property is not necessarily in
> the PHY node either.
> 
> > > Simply leaving the refclk always on in the PHY seems like a bit of a
> > > hack and I'm not even sure that can be considered correct.
> > 
> > I wouldn't agree it is a hack, even though it may sound like one. The option to
> > keep refclk always on in the PHY is precisely there for usecase like this.
> 
> I just skimmed the spec so perhaps I'm missing something, but there's
> definitely wordings in there that explicitly says that L1 PM substates
> must not be enabling unless you have a functioning CLKREQ# signal.
> 

I checked 'PCI Express Base spec 4.0', and there is a wording in 'Section 5.5
L1 PM Substates':

For L1 PM ... each port must have a unique instance of the signal (CLKREQ#) and
the upstream and downstream port CLKREQ# signals must be connected.

So yes, we should not enable L1 PM substates in this case. And also it appears
to me that the property should be part of the controller node, since CLKREQ# is
handled by the controller instance.

- Mani

> Johan

-- 
மணிவண்ணன் சதாசிவம்