[PATCH 0/5] PM: Enable option of re-use runtime PM callbacks at system suspend

Ulf Hansson ulf.hansson at linaro.org
Fri Nov 29 04:35:57 EST 2013


On 29 November 2013 10:32, Ulf Hansson <ulf.hansson at linaro.org> wrote:
>>
>> The lack of specificity here doesn't make the discussion any easier.
>>
>> It usually is better to talk about specific problems to address than
>> using general terms which may mean slightly different things for different
>> people.
>
> During these discussions, I have tried to point at existing code for
> drivers and existing code for power domains. Those which at the moment
> either have got things wrong or are unnecessary complicated, in
> regards to their PM implementation.
>
> I suppose I could provide some more patches for proof of concept, will
> that be a way forward?
>
>>
>>> >> Additionally, some drivers seems to have messed up things when combining
>>> >> runtime PM with system PM. While we enable the option of re-using the runtime
>>> >> PM callbacks during system PM, we also intend to clarify the way forward for
>>> >> how these scenarios could be resolved.
>>> >>
>>> >> Some new helper macros for defining PM callbacks and two new pm_generic*
>>> >> functions has been implemented in this patch set. These are provided to make it
>>> >> easier for those who wants to enable the option of re-using the runtime PM
>>> >> callbacks during system suspend.
>>> >
>>> > I'm generally opposed to re-using callbacks like this, because it adds confusion
>>> > to the picture.  It may seem to be clever, but in fact it leads to bad design
>>> > choices in the drivers in my opinion.
>>>
>>> In my world of the kernel, it will clearly resolve confusions and
>>> simplify a significant amount of code in power domains, buses and
>>> drivers. So I guess it depends on from what point you look at this.
>>
>> This is so vague that I don't even know how to respond. :-)
>>
>> So let me say instead that what you did in patch [5/5] is a layering violation
>> which always is a bug, even if it doesn't break things outright.
>>
>> After that patch the driver would call into a layer that is supposed to call
>> into it under normal conditions.  Moreover, it would expect that layer to
>> call back into it again in a specific way, which may or may not happen depending
>> on how exactly that layer is implemented.  So even if it works today, it will
>> add constraints on how that other layer may be implmented which *is* confusing
>> and wrong in my not so humble opinion.
>>
>> I'll never apply any patches that lead to situations like that, don't even
>> bother to send them to me.  Pretty please.
>>
>
> After all these good discussions which clearly pointed the solution
> into this direction, you decide to bring up this argument now? It
> makes me wonder.
>
> Indirectly what you are saying is that, the PM core should at
> device_prepare, do pm_runtime_disable() instead of
> pm_runtime_get_noresume(), to prevent drivers/subsystems from
> triggering "layering violations" by invoking pm_runtime_get_sync().
>
> Because, this is exactly the same kind of layering violation you refer
> to while neglecting my approach, which at the moment
> drivers/subsystems not only are allowed to, but also encouraged to do
> during system suspend.
>
> Now, obviously I don't think we shall change the behaviour of PM core,
> that would just not be convenient for subsystems and drivers, right?
>
> So, the PM core allows layering violations for the .runtime_resume
> callbacks to be invoked during system suspend. It can do so, because
> it trust drivers/subsystems to act responsibly and to what suites them
> best.
>
> For the same reasons, I believe we should trust drivers/subsystems, to
> understand when it makes sense for them to re-use all of the runtime
> PM callbacks during system suspend and not just the .runtime_suspend
> callback.

Sorry, another typo:

"not just the .runtime_suspend" -> "not just the .runtime_resume".

>
> That is in principle what I and Alan, who came up with this idea, are
> suggesting.
>
>>> And, as you stated previously during these discussions, we have the
>>> opportunity to update the documentation around this topic, I will
>>> happily do it, if needed.
>>
>> That's always welcome. :-)
>>
>>> >
>>> > Let's talk about specific examples, though.
>>> >
>>> > Why exactly do you need what patch [5/5] does in the exynos_drm_fimc driver?
>>>
>>> This was a simple example, I wanted to visualize how the new building
>>> blocks were going to be used. Anyway, this we achieve with the patch:
>>>
>>> 1.
>>> The PM part in the driver becomes simplified, we don't need the
>>> wrapper functions for the runtime PM callbacks any more.
>>
>> No, it is not simplified.  It becomes *far* more complicated conceptually
>> instead, although that is hidden by moving the complexity into the functions
>> added by patch [1/5].  So whoever doesn't look into those functions will
>> not actually realize how complicated the code really is.
>>
>>> 2.
>>> Previously the driver did not make sure runtime PM was disabled,
>>> before it put the device into low power state at .suspend. From a
>>> runtime PM point of view, this is not a "nice" behaviour and opens up
>>> for confusions, even if it likely would work in most cases.
>>
>> So the proper fix, in my opinion, would be to point .suspend_late and
>> .resume_early in that driver to fimc_suspend() and fimc_resume(),
>> respectively, and leave the .suspend and .resume pointers unpopulated.
>>
>> Wouldn't that actually work?
>
> If we decide to ignore the power domain issue below, yes.
>
>>
>>> 3.
>>> The power domain runtime PM callbacks were by-passed during system
>>> suspend, my patch fixes this.
>>
>> I don't exactly understand this.  Why would they be bypassed?
>>
>>> Why do I want this? Because the power
>>> domain can have runtime PM resources it need to handle at this phase.
>>> Potentially, it could handle that from it's .suspend_late callback
>>> instead, but then it gets unnecessary complicated, which is what
>>> clearly happened to the power domain for OMAP2, for example.
>>
>> I'd like to understand this.  What exactly do you mean by "unnecessary
>> complicated"?
>
> Please, have a deeper look into the OMAP power domain implementation.
>
> If each SoC (for those that have power domain regulators) needs their
> own version of such a power domain, then I certainly think it is more
> complicated that in needs to be.
>
>>
>>>
>>> If you want additional proof of concepts, we can have a look at more
>>> complex example.
>>>
>>> Typically I am thinking of cases were a cross SoC driver is attached
>>> to a bus and for some SoCs a power domain as well. Then, the bus, the
>>> power domain and the driver - all have runtime PM resources to handle.
>>
>> Sure.
>
> OK, I consider resending the patch set, including some additional
> proof of concept patches.
>
>>
>>> In these cases using the new building blocks will not only
>>> significantly simplify code, but also fix immediate bugs. One example
>>> are drivers attached to the AMBA bus, at drivers/amba/bus.c.
>>
>> Again, what drivers and what's the bug you're talking about?
>
> I will use some of these as examples, it will be more visible to you then.
>
> Kind regards
> Uffe
>
>>
>> Rafael
>>



More information about the linux-arm-kernel mailing list