Re: [PATCH 0/2] cpufreq/opp: rework regulator initialization

[Rafael J. Wysocki]

On Fri, Feb 8, 2019 at 1:04 PM Marek Szyprowski
<m.szyprowski@xxxxxxxxxxx> wrote:
>
> Hi Sudeep,
>
> On 2019-02-08 12:51, Sudeep Holla wrote:
> > On Fri, Feb 08, 2019 at 12:47:06PM +0100, Marek Szyprowski wrote:
> >> On 2019-02-08 12:00, Sudeep Holla wrote:
> >>> On Thu, Feb 07, 2019 at 01:22:25PM +0100, Marek Szyprowski wrote:
> >>>> Dear All,
> >>>>
> >>>> This is a scenario that triggers the above issue:
> >>> [...]
> >>>> 1. system disables non-boot cpu's at the end of system suspend procedure,
> >>>> 2. this in turn deinitializes cpufreq drivers for the disabled cpus,
> >>>> 3. early in the system resume procedure all cpus are got back to online
> >>>>    state,
> >>>> 4. this in turn causes cpufreq to be initialized for the newly onlined
> >>>>    cpus,
> >>>> 5. cpufreq-dt acquires all its resources (clocks, regulators) during
> >>>>    ->init() callback,
> >>> This is strictly not just restricted to cpufreq-dt, but to any driver
> >>> supporting multiple policies. So we need a generic fix not just
> >>> cpufreq-dt specific.
> >> Could you point which other driver needs similar fix? Here in cpufreq-dt
> >> the problem was caused by using regulator api (indirectly) from
> >> ->init(). All other drivers, which have regulators support, are for old,
> >> obsolete, uni-processor systems, which don't have the problem of
> >> secondary cpu suspend during system suspend/resume cycle.
> >>
> > scmi_cpufreq for instance. We can fix that in driver my moving to polling
> > to get cpufreq_get_rate, but we support both polling and interrupt based.
> > We may wait for remote processor interrupt in get_rate.
>
> Frankly, I don't feel I know enough to touch this driver and I don't
> think that this can even be fixed in a generic way in the cpufreq core.
> Maybe a comment somewhere is needed that ->init() might be called during
> early system resume with irqs off and driver is responsible for handling
> such case until the proper resume?

Well, adding a comment to that effect certainly won't hurt as that's
how things work now.

However, it looks like something needs to be done in addition to that.