Re: [PATCH] PM / runtime: Drop children check from __pm_runtime_set_status()

From: Ulf Hansson
Date: Mon Dec 04 2017 - 05:41:30 EST


On 1 December 2017 at 12:03, Yoshihiro Shimoda
<yoshihiro.shimoda.uh@xxxxxxxxxxx> wrote:
> Hi,
>
>> From: Ulf Hansson, Sent: Friday, December 1, 2017 6:22 PM
>>
>> + Kishon
>>
>> On 30 November 2017 at 13:51, Yoshihiro Shimoda
>> <yoshihiro.shimoda.uh@xxxxxxxxxxx> wrote:
>> > Hi,
>> >
>> >> From: Ulf Hansson, Sent: Wednesday, November 29, 2017 6:59 PM
>> >>
>> >> On 29 November 2017 at 10:43, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>> >> > Hi Ulf,
>> > <snip>
>> >> Okay, so the problem remains no matter which solution for wakeup you
>> >> pick in genpd.
>> >
>> > Yes. Today I could reproduce this issue without usb host driver.
>> > - The renesas_usb3 usb peripheral driver has generic phy handling.
>> > (The peripheral driver uses different generic phy driver (phy-rcar-gen3-usb3.c) though.)
>> > --> If I used the current renesas_usb3 (this means doesn't call phy_power_{on,off}(),
>> > the issue didn't happen.
>> > --> If I added phy_power_{on,off}() calling, the issue happened.
>> > --> So, I'm thinking the APIs are related to the issue.
>>
>> Yes.
>>
>> >
>> > - The generic phy APIs are in drivers/phy/phy-core.c.
>> > --> The phy-rcar-gen3-usb[23] drivers call only pm_runtime_enable() before devm_phy_create().
>> > --> The phy-core will call pm_runtime_{get_sync,put}() in phy_{init,exit,power_{on,off}}.
>> > --> So, IIUC, both devices of phy-<dev_name>.<id> and <dev_name> will be handled by runtime PM APIs.
>> > --> The runtime PM implementation of phy-core seems good to me. But...?
>>
>>
>> I have digested the information that you and Geert provided, thanks!
>>
>> So, my conclusions so far is:
>>
>> The phy core is using runtime PM reference counting at
>> phy_power_on|off(). Although it does that on the phy core device,
>> which is a child device of the phy provider device.
>>
>> Because phy_power_off() is called during system suspend from phy
>> consumer drivers like usb, the phy core device (child) and the phy
>> provider device (parent) will never become runtime suspended (because
>> the PM core has invoked pm_runtime_get_no_resume() for all device in
>> the device prepare phase).
>>
>> Then, when genpd calls pm_runtime_force_suspend() at the suspend noirq
>> phase for the phy provider device, the call to
>> pm_runtime_set_suspended() in there, triggers the earlier error
>> message, which is because the child (phy core device) is still runtime
>> resumed.
>
> Thank you very much for the conclusions!
> It's helpful to me about runtime PM behavior.
>
>> >> Then this seems to point to that the driver may be misbehaving in some
>> >> way. I can help to check what is going on.
>> >
>> > I guess so. But, I don't find yet...
>>
>> I think the below patch will help, although I am not sure if that is
>> sufficient as a long term fix.
>
> Thank you very much for your help!
> Also, I'm not sure how to fix for a long term kernels though...
>
>> Can you please try and see if it solves the problems?
>
> Sure! I tested your patch, and then the following message disappeared!
>
> Enabling runtime PM for inactive device (ee080200.usb-phy) with active children

Great, that confirms my theory.

I will re-work the patch and re-post it to see what people thinks about it.

>
> However, the following message still exists.
>
> Enabling runtime PM for inactive device (ee080000.usb) with active children
>
> So, I guess ohci-platform.c also has similar issue.

Yes, very likely!

However, I need some more time to look into this to be able to suggest
a solution.

>
> JFYI, the ehci-platform.c doesn't have runtime PM handling.
> So, I think that error message doesn't output from ehci devices.

Right, thanks!

Kind regards
Uffe