Re: [PATCH v5 1/2] usb: host: plat: Enable xhci plat runtime PM

[Baolin Wang]

Hi Mathias,

On 21 February 2017 at 10:09, Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
>>>>>> Do you have any comments about this patch? Thanks.
>>>>>>
>>>>>>>    drivers/usb/host/xhci-plat.c |   41
>>>>>>> ++++++++++++++++++++++++++++++++++++-----
>>>>>>>    1 file changed, 36 insertions(+), 5 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/usb/host/xhci-plat.c
>>>>>>> b/drivers/usb/host/xhci-plat.c
>>>>>>> index ed56bf9..5805c6a 100644
>>>>>>> --- a/drivers/usb/host/xhci-plat.c
>>>>>>> +++ b/drivers/usb/host/xhci-plat.c
>>>>>>> @@ -246,6 +246,10 @@ static int xhci_plat_probe(struct platform_device
>>>>>>> *pdev)
>>>>>>>           if (ret)
>>>>>>>                   goto dealloc_usb2_hcd;
>>>>>>>
>>>>>>> +       pm_runtime_get_noresume(&pdev->dev);
>>>>>>> +       pm_runtime_set_active(&pdev->dev);
>>>>>>> +       pm_runtime_enable(&pdev->dev);
>>>>>>> +
>>>>>>>           return 0;
>>>>>>>
>>>>>>>
>>>>>
>>>>> To me this looks like we are not really enabling runtime pm for xhci, we
>>>>> increment the
>>>>> usage count in probe, and keep it until remove is called.
>>>>>
>>>>> This would normally prevent parent dwc3 from runtime suspending, but I
>>>>> see
>>>>> that in patch 2/2 adds
>>>>> pm_suspend_ignore_children(dev, true) to dwc3, and, that dwc3 actually
>>>>> controls xhci runtime
>>>>> pm by calling pm_runtime_get/put for xhci in its own dwc3
>>>>> runtime_suspend/resume callbacks.
>>>>>
>>>>> Looks a bit upside down to me, what's the reason for this?
>>>>>
>>>>> what prevents calling pm_runtime_put_* before leaving probe in xhci and
>>>>> let
>>>>> pm code handle
>>>>> the runtime suspend and parent/child relationships?
>>>>
>>>>
>>>> When dwc3 controller is working on host mode, which will create and
>>>> add the USB hcd for host side. Then if we want to suspend dwc3
>>>> controller as host mode, the USB device (bus) of host side will
>>>> runtime suspend automatically if there are no slave attached (by
>>>>
>>>> usb_runtime_suspend()--->usb_suspend_both()--->usb_suspend_interface()--->usb_suspend_device()--->generic_suspend()--->hcd_bus_suspend()--->xhci_bus_suspend()),
>>>> but we should not suspend xHCI device (issuing xhci_suspend()) now,
>>>> since some other  controllers did not implement the runtime PM to
>>>> control xHCI device's runtime state, which will cause other
>>>> controllers can not resume xHCI device (issuing xhci_resume()) if the
>>>> xHCI device suspend automatically by its child device. Thus we should
>>>> get the runtime count for xHCI device in xhci_plat_probe() in case
>>>> xHCI device will also suspend automatically by its child device.
>>>> According to that, for xHCI device's parent: dwc3 device, we should
>>>> put the xHCI device's runtime count to suspend xHCI device manually.
>>>>
>>>
>>> Any more comments?
>>>
>>
>> I think I at least partially understand your point. You don't want to enable
>> runtime suspend for all xhci platform devices by default, but only for the
>> ones
>> that are part of dwc3.
>>
>> The implementation you suggest is that xhci platform driver always increase
>> the xhci platform
>> device usage counter during probe with pm_runtime_get_noresume(), and never
>> decrement it,
>> preventing xhci platform devices from runtime suspending by themselves.
>>
>> You then control xhci runtime suspend from dwc3 driver runtime suspend,
>> allowing
>> xhci platfrom controller to runtime suspend only when dwc3 runtime suspends
>> by decrementing xhci
>> platform device usage in dwc3 runtime suspend callback.
>> As xhci is a child of dwc3 in this case,  dwc3 would never runtime suspend
>> as long as xhci is running, so
>> dwc3 needed to be detached from xhci by setting dwc3 to ignore its children
>> to get it to work.
>
> Yes, that's my point.
>
>>
>> I don't yet understand why we can't just keep runtime pm disabled as a
>> default for xhci platform devices.
>> It could be enabled by whatever creates the platform device by setting some
>> device property
>> (or equivalent), which would be checked in xhci_plat_probe() before enabling
>> runtime pm. It
>> could then optionally be set in sysfs via power/control entry.
>
> I think runtime pm is not one hardware property, is it suitable if we
> introduce one device property to enable/disable runtime pm?

As I said, runtime pm is not one hardware property, I think it is not
suitable if we introduce one device property to enable/disable runtime
pm.

Secondly, we only can resume the xhci platform device by getting the
xhci usage counter from gadget driver, since the cable plug in/out
events only can be notified to glue layer of gadget driver(like dwc3
glue layer). That means if we want to suspend xhci platform device, we
must put xhci usage counter (so we can not keep their parent-child
relationship intact). That is why we need
pm_suspend_ignore_children(dev, true).

Mathias, could you re-check this patch? Thanks.

>>
>> We would keep the usage counting intact, and only enable xhci platform
>> device runtime pm
>> if the platform supports it, and dwc3 and xhci would keep their parent-child
>> relationship
>> intact and make sure dwc3 can't runtime suspend before xhci.
>
> Yes, that is the benefits.
>
>>
>> To the concerns about xhci runtime suspending too early, the codepath you
>> describet only applies
>> for roothub devices, which sohuld be called only after other devices have
>> suspended.
>>
>> To make sure xhci platform device does not runtime suspend during probe
>> after first
>> hcd is added (without children), but before second one, make sure to
>> increase the usage count before creating and adding
>> both hcd's, and only decrease it after both are  ready.
>>
>> This way it will only runtime suspend after both buses are created.
>>
>> xhci_plat_probe(pdev)
>>   /* prevent platform device runtime suspend between adding main and
>> shared_hcd */
>>   pm_runtime_get_noresume(&pdev->dev)
>>   usb_create_hcd()
>>   usb_create_shared_hcd()
>>   usb_add_hcd(hcd)
>>   usb_add_hcd(shared_hcd)
>>   /* both hcs's added, allow platform device to runtime suspend */
>>   pm_runtime_put_noidle(&dev->dev);
>>
>
> Make sense.
>
> --
> Baolin.wang
> Best Regards



-- 
Baolin.wang
Best Regards