Re: [PATCH 1/3] phy: core: Move runtime PM reference counting to the parent device

[Ulf Hansson]

On 20 December 2017 at 10:02, Kishon Vijay Abraham I <kishon@xxxxxx> wrote:
> Hi Ulf,
>
> On Wednesday 20 December 2017 02:05 PM, Ulf Hansson wrote:
>> On 20 December 2017 at 07:42, Kishon Vijay Abraham I <kishon@xxxxxx> wrote:
>>> Hi Ulf,
>>>
>>> On Wednesday 20 December 2017 02:52 AM, Ulf Hansson wrote:
>>>> The runtime PM deployment in the phy core is a bit unnecessary complicated
>>>> and the main reason is because it operates on the phy device, which is
>>>> created by the phy core and assigned as a child device of the phy provider
>>>> device.
>>>>
>>>> Let's simplify the code, by replacing the existing calls to
>>>> phy_pm_runtime_get_sync() and phy_pm_runtime_put(), with regular calls to
>>>> pm_runtime_get_sync() and pm_runtime_put(). While doing that, let's also
>>>> change to give the phy provider device as the parameter to the runtime PM
>>>> calls. This together with adding error paths, that allows the phy
>>>> provider device to be runtime PM disabled, enables further clean up the
>>>> code. More precisely, we can simply avoid to enable runtime PM for the phy
>>>> device altogether, so let's do that as well.
>>>>
>>>> More importantly, this change also fixes an issue for system suspend.
>>>> Especially in those cases when the phy provider device gets put into a low
>>>> power state via calling the pm_runtime_force_suspend() helper, as is the
>>>> case for a Renesas SoC, which has the phy provider device attached to the
>>>> generic PM domain.
>>>>
>>>> The problem in this case, is that pm_runtime_force_suspend() expects the
>>>> child device of the provider device to be runtime suspended, else this will
>>>> trigger a WARN splat (correctly) when runtime PM gets re-enabled at system
>>>> resume.
>>>>
>>>> In the current case, even if phy_power_off() triggers a pm_runtime_put()
>>>> during system suspend the phy device (child) doesn't get runtime suspended,
>>>> because that is prevented in the system suspend phases. However, by
>>>> avoiding to enable runtime PM, this problem goes away.
>>>>
>>>> Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
>>>> ---
>>>>  drivers/phy/phy-core.c | 33 +++++++++++++--------------------
>>>>  1 file changed, 13 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/phy/phy-core.c b/drivers/phy/phy-core.c
>>>> index b4964b0..9fa3f13 100644
>>>> --- a/drivers/phy/phy-core.c
>>>> +++ b/drivers/phy/phy-core.c
>>>> @@ -222,10 +222,10 @@ int phy_init(struct phy *phy)
>>>>       if (!phy)
>>>>               return 0;
>>>>
>>>> -     ret = phy_pm_runtime_get_sync(phy);
>>>> -     if (ret < 0 && ret != -ENOTSUPP)
>>>> +     ret = pm_runtime_get_sync(phy->dev.parent);
>>>
>>> Won't this make phy-core manage pm_runtime of phy_provider even though the
>>> phy_provider might not intend it?
>>
>> No it shouldn't.
>>
>> There are two cases to consider around this.
>>
>> 1) CONFIG_PM is unset. In this case pm_runtime_get_sync() will return
>> 1, which is treated as succeeds by the error path.
>>
>> 2) CONFIG_PM is set, but the phy provider don't use runtime PM, thus
>> it hasn't called pm_runtime_enable() for its device. In this case,
>> pm_runtime_get_sync() returns -EACCES, which is also treated as
>> success by the error path.
>
> There can be a case where the phy_provider uses runtime PM but doesn't want
> phy-core to manage it.

Ah, so you mean there are cases when the provider driver calls
pm_runtime_enable() *after* it calls phy_create()/dev_phy_create()
instead of before?

I am not really sure I understand *why* a provider driver wants to do
that though, do you have more details?
I mean, even if the phy core handles runtime PM, additional management
can be done on top in the phy provider, there is nothing preventing
that, but I guess that isn't sufficient?

Kind regards
Uffe