Re: [RFC PATCH 1/2] genirq: Add runtime resume/suspend support for IRQ chips

Re: [RFC PATCH 1/2] genirq: Add runtime resume/suspend support for IRQ chips

[Geert Uytterhoeven]

Hi Jon,

On Mon, Nov 16, 2015 at 11:34 AM, Jon Hunter <jonathanh-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org> wrote:
> On 16/11/15 09:49, Geert Uytterhoeven wrote:
>> On Mon, Nov 16, 2015 at 10:46 AM, Jon Hunter <jonathanh-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org> wrote:
>>> On 13/11/15 20:01, Thomas Gleixner wrote:
>>>> On Fri, 13 Nov 2015, Jon Hunter wrote:
>>>>> On 12/11/15 23:20, Kevin Hilman wrote:
>>>>>> If all the RPM devices in the domain go idle, it will be powered off
>>>>>> independently of the status of the irqchip because the irqchip isn't
>>>>>> using RPM.
>>>>>
>>>>> That's dependent on how the irqchip uses these helpers. If these helpers
>>>>> invoke RPM then that will not be the case.
>>>>
>>>> You need a very proper description of how that domain is working. If
>>>> all devices are idle, it's not necessary correct to power down the
>>>> irqchip as is might serve other devices as well.
>>>
>>> Agreed. The irqchip should only be powered down if there are no
>>> interrupts in-use/requested. Runtime-pm will keep a reference count for
>>> all requested IRQs.
>>
>> That means the irqchip won't be powered down automatically when the
>> last user is powered down, unless all users release their irqs during
>> suspend.
>
> Right.
>
>> Handling it automatically needs more bookkeeping than a simple reference
>> count.
>
> So what would you suggest? Adding a pm_runtime_register_irq() API that
> would register an IRQ with the device that you want RPM to handle? Not
> sure if there is a better/easier way to handle this.

The irqchip needs to keep track how many times request_irq() has been
called, cfr. your suggestion above.

On the other side, the system needs to keep track how many times request_irq()
has been called for each irqchip, so it can subtract those numbers from the
irqchip's counters during suspend of the device, and re-add them during resume.
So we need at least a "struct device *" parameter for request_irq().
devm_request_irq() already has that, but not all drivers use that.

However, I think this should be looked at into the context of "[RFD]
Functional dependencies between devices".
https://lwn.net/Articles/662205/
https://lkml.org/lkml/2015/10/27/388

There can be other dependencies than interrupts between devices.
All functions using dependencies need a "struct device *" parameter to
record information.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert-Td1EMuHUCqxL1ZNQvxDV9g@public.gmane.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Re: [RFC PATCH 1/2] genirq: Add runtime resume/suspend support for IRQ chips

[Jon Hunter]

Hi Geert,

On 16/11/15 10:48, Geert Uytterhoeven wrote:
> On Mon, Nov 16, 2015 at 11:34 AM, Jon Hunter <jonathanh-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org> wrote:

[snip]

>>> Handling it automatically needs more bookkeeping than a simple reference
>>> count.
>>
>> So what would you suggest? Adding a pm_runtime_register_irq() API that
>> would register an IRQ with the device that you want RPM to handle? Not
>> sure if there is a better/easier way to handle this.
> 
> The irqchip needs to keep track how many times request_irq() has been
> called, cfr. your suggestion above.
> 
> On the other side, the system needs to keep track how many times request_irq()
> has been called for each irqchip, so it can subtract those numbers from the
> irqchip's counters during suspend of the device, and re-add them during resume.
> So we need at least a "struct device *" parameter for request_irq().
> devm_request_irq() already has that, but not all drivers use that.

Yes that would make sense. However, I am wondering if the
syscore suspend/resume operators could be used here to do something
like ...

	pm_runtime_disable(dev);
	if (!pm_runtime_status_suspended(dev))
		chip->irq_runtime_suspend(data);

> However, I think this should be looked at into the context of "[RFD]
> Functional dependencies between devices".
> https://lwn.net/Articles/662205/
> https://lkml.org/lkml/2015/10/27/388
> 
> There can be other dependencies than interrupts between devices.
> All functions using dependencies need a "struct device *" parameter to
> record information.

Yes I like the sound of that. That would be ideal. However, I am
guessing that that is a way off at the moment ...

Cheers
Jon