Re: PM/runtime issue: Device cannot runtime resume because power domain is off during system resume

[Krzysztof Kozlowski <k.kozlowski@samsung.com>]

On śro, 2014-10-29 at 14:01 +0100, Ulf Hansson wrote:
> On 16 October 2014 16:05, Krzysztof Kozlowski <k.kozlowski@samsung.com> wrote:
> > Hi,
> >
> > I encounter issues with DRM/panel driver for Exynos after resuming from
> > suspend to RAM. This is reproduced on 3.16 but I think it is applicable
> > also to mainline.
> >
> > The scenario:
> > 0. DRM DSI is in LCD power domain. Power domain is OFF before
> > suspending.
> > 1. System suspends.
> > 2. System resumes.
> > 3. DRM driver:resume().
> > 4. pm_runtime_get() for DRM DSI driver.
> > 5. This should enable LCD power domain.
> > 6. but __pm_genpd_poweron() exits early because:
> >    genpd->prepared_count = 3
> 
> I am getting to this soonish in my genpd work. Currently I am working
> on the boot issues. :-)
> 
> At a first glance, I think the genpd->prepare_count variable shouldn't
> exist at all, but I need to dig into this in more detail to be sure.
> 
> In principle I think genpd should rely _more_ on the PM core during
> system PM transitions. For example the PM core invokes a
> pm_runtime_get_noresume() (and for some reason genpd as well) at the
> ->prepare() phase. Shouldn't that be enough!?
> 
> >    genpd->suspend_power_off = 1
> >   in domain.c:183
> >
> > 7. The domain is not powered on but it is marked as active.
> > 8. The DRM driver thinks everything is OK and proceeds with resume...
> > but it fails because whole power domain is OFF.
> >
> > The question: Shouldn't the power domain be powered up EARLY?
> 
> Good question, I would like us to strive towards these goals:
> 
> For CONFIG_PM_RUNTIME, I think the PM domain should be powered up at
> the first time a runtime PM resume callbacks gets invoked.
> For !CONFIG_PM_RUNTIME, I think we should restore the state the PM
> domain had prior we went to system PM sleep.

For the first case (PM_RUNTIME) I sent a proposal:
https://lkml.org/lkml/2014/10/23/308
which seems to work for my configuration. However I did not test it
without runtime PM...

The issues I found were:
1. Runtime resume during early system suspend (triggered by poking
console).
2. Runtime resume during system resume (triggered by "normal" resuming
device).
In both case the power domain was off so runtime resume failed.

I've got a simple testing scenario and setup for reproducing these so if
you need tests I can help. 

Best regards,
Krzysztof