Re: [PATCH V2 1/2] opp: Allow dev_pm_opp_get_opp_table() to return -EPROBE_DEFER

[Sudeep Holla]

On Mon, Oct 19, 2020 at 02:54:11PM +0530, Viresh Kumar wrote:
> On 19-10-20, 10:17, Sudeep Holla wrote:
> > On Mon, Oct 19, 2020 at 10:28:27AM +0530, Viresh Kumar wrote:
> > > On 16-10-20, 12:12, Sudeep Holla wrote:
> > > > On Fri, Oct 16, 2020 at 07:00:21AM +0100, Sudeep Holla wrote:
> > > > > On Fri, Oct 16, 2020 at 09:54:34AM +0530, Viresh Kumar wrote:
> > > > > > On 15-10-20, 19:05, Sudeep Holla wrote:
> > > > > > > OK, this breaks with SCMI which doesn't provide clocks but manage OPPs
> > > > > > > directly. Before this change clk_get(dev..) was allowed to fail and
> > > > > > > --EPROBE_DEFER was not an error.
> > > > > >
> > > > > > I think the change in itself is fine. We should be returning from
> > > > > > there if we get EPROBE_DEFER. The question is rather why are you
> > > > > > getting EPROBE_DEFER here ?
> > > > > >
> > > > >
> > > > > Ah OK, I didn't spend too much time, saw -EPROBE_DEFER, just reverted
> > > > > this patch and it worked. I need to check it in detail yet.
> > > > >
> > > > 
> > > > You confused me earlier. As I said there will be no clock provider
> > > > registered for SCMI CPU/Dev DVFS.
> > > > 	opp_table->clk = clk_get(dev, NULL);
> > > > will always return -EPROBE_DEFER as there is no clock provider for dev.
> > > > But this change now propagates that error to caller of dev_pm_opp_add
> > > > which means we can't add opp to a device if there are no clock providers.
> > > > This breaks for DVFS which don't operate separately with clocks and
> > > > regulators.
> > >
> > > The CPUs DT node shouldn't have a clock property in such a case and I
> > > would expect an error instead of EPROBE_DEFER then. Isn't it ?
> > 
> > Ideally yes, but for legacy reasons clocks property has been used for
> > providing OPP/DVFS handle too. While we can change and add new property
> > for that, it will still break old bindings.
> 
> I am not sure I understood it all. So does your platform have the
> clock-names property or not for the CPUs ? And how will something
> break here ?
> 

Yes it has clocks property but used by SCMI(for CPUFreq/DevFreq) and not
by any clock provider driver. E.g. the issue you will see if "clocks"
property is used instead of "qcom,freq-domain" on Qcom parts.

On SCMI, we have used clocks property to represent perf domain which I
understand is not ideal but it is there 🙁.

-- 
Regards,
Sudeep