[PATCH v2 1/9] i2c: prepare runtime PM support for I2C client devices

[Sylwester Nawrocki]

On 09/13/2013 05:40 PM, Mika Westerberg wrote:
[...]
>>> The call to pm_runtime_get_noresume() should make sure that the device is
>>> in active state (at least in state where it can access the bus) if I'm
>>> understanding this right.
>>
>> I can't see how this would happen. How runtime_resume/runtime_suspend
>> callbacks would get invoked with this code, if, e.g. originally driver called
>> pm_runtime_enable(), pm_runtime_get_sync(), pm_runtime_put_sync() in probe() ?
>
> The driver callbacks are not called but if the device has been attached to
> a power domain (like we do with ACPI) the power domain callbacks get called
> and it brings the "bus" to such state that we are able to access the
> device. That also was the reason I used _noresume() but didn't look too
> close the implementation.

OK, but if a client driver assumes default inactive power state it will 
expect
its callbacks to get called. Otherwise exisiting code might break. So, e.g.
in case of s5p-tv it would rather need to be something like:

	pm_runtime_put()

	pm_runtime_get_sync()
  	sii9234_verify_version()
  	pm_runtime_put(dev)

>> pm_runtime_get_noresume() merely increments usage counter of a device.
>> It seems that these changes will break the s5p-tv driver. I might be missing
>> something though.
>
> You are right and Kevin also mentioned this. It should be pm_runtime_get(),
> if I'm not mistaken.

Note that client drivers usually call pm_runtime_enable() only when it 
is safe
to call their driver's runtime PM callbacks. By enabling runtime PM 
before the
client's driver has completely initialized we may risk that the 
callbacks are
executed with uninitialized data, if I understand things correctly.

>> As Mark pointed out this is currently unwanted behaviour to runtime PM
>> activate a bus controller device manually in the core for when the client's
>> probe() is executed, since i2c core will activate the bus controller for when
>> transfer is being carried out.
>>
>> But I can understand this is needed for ACPI and it shouldn't break existing
>> drivers, that do runtime PM activate the client device in probe().
>
> Indeed, we don't want to break anything (and we still need something like
> this for ACPI).
>
>> Now I'm sure this will break power management of the drivers/media/exynos4-is
>> driver, due to incorrect power sequence (power domain and clocks handling).
>> I'll try to take care of it in separate patch, as I have some patches pending,
>> that move most of code from drivers/media/exynos4-is/fimc-is-sensor.c to
>> drivers/media/i2c/s5k6a3.c.
>
> I missed that code when I converted existing users to this method. Sorry
> about that (I can handle that in the next version).
>
> I quickly looked at it and I don't see anything that could break (once
> converted). What it does is this:
>
> 	pm_runtime_no_callbacks(dev);
> 	pm_runtime_enable(dev);
>
> changing that to:
>
> 	pm_runtime_no_callbacks(dev);
> 	pm_runtime_put(dev);
>
> shouldn't cause problems AFAICT.

Yes, considering this driver in isolation it should be fine.

However, I observed system suspend issues when the I2C bus controller was
being activated (which would happen in the I2C core after your changes)
before some other driver has initialized.

So to ensure things continue to work the "fimc-isp-i2c" driver would need
to be registered after the "exynos4-fimc-is" driver has initialized. Or the
"exynos4-fimc-is" would need to call of_platform_populate() to instantiate
its all children devices as specified in device tree (see arch/arm/boot/dts/
exynos4x12.dtsi in -next). "simple-bus" would then have to be not listed in
the compatible property of that top level device. So to avoid regressions
some additional changes would be needed, outside of this particular I2C
client driver. I guess this could be avoided by better design of the
exynos4-is driver right from the beginning. But it's all some times tricky
when there is some many IP blocks involved and the hardware behaviour/device
interactions are not always well documented.