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

[Jon Hunter]

On 12/11/15 13:47, Lars-Peter Clausen wrote:
> On 11/12/2015 02:35 PM, Jon Hunter wrote:
>>
>> On 12/11/15 13:20, Lars-Peter Clausen wrote:
>>> On 11/12/2015 11:59 AM, Jon Hunter wrote:
>>>>
>>>> On 11/11/15 15:41, Grygorii Strashko wrote:
>>>>> On 11/11/2015 12:13 PM, Jon Hunter wrote:
>>>>>>
>>>>>> On 10/11/15 18:07, Lars-Peter Clausen wrote:
>>>>>>> On 11/10/2015 05:47 PM, Grygorii Strashko wrote:
>>>>>>> [...]
>>>>>>>>> I was trying to simplify matters by placing the resume call in
>>>>>>>>> __setup_irq() as opposed to requested_threaded_irq(). However, the would
>>>>>>>>> mean the resume is inside the bus_lock and may be I should not assume
>>>>>>>>> that I can sleep here.
>>>>>>>>>
>>>>>>>>>> Can you folks please agree on something which is correct and complete?
>>>>>>>>>
>>>>>>>>> Soren I am happy to defer to your patch and drop this. My only comment
>>>>>>>>> would be what about the request_percpu_irq() path in your patch?
>>>>>>>>>
>>>>>>>>
>>>>>>>> I have the same comment here as I asked Soren:
>>>>>>>> 1) There are no restrictions to call irq set_irq_type() whenever,
>>>>>>>> as result HW can be accessed before request_x_irq()/__setup_irq().
>>>>>>>> And this is used quite widely now :(
>>>>>>>>
>>>>>>>
>>>>>>> Changing the configuration of a resource that is not owned seems to be
>>>>>>> fairly broken. In the worst case this will overwrite the configuration that
>>>>>>> was set by owner of the resource.
>>>>>>>
>>>>>>> Especially those that call irq_set_irq_type() directly before request_irq(),
>>>>>>> given that you supply the trigger type to request_irq() which will make sure
>>>>>>> that there are no conflicts and the configure.
>>>>>>>
>>>>>>> This is a bit like calling gpio_set_direction() before you call
>>>>>>> gpio_request(), which will also have PM issues.
>>>>>>
>>>>>> Yes, I agree that this does sound a bit odd, but ...
>>>>>>
>>>>>>>> For example, during OF boot:
>>>>>>>>
>>>>>>>> [a]  irq_create_of_mapping()
>>>>>>>>     - irq_create_fwspec_mapping()
>>>>>>>>       - irq_set_irq_type()
>>>>>>
>>>>>> The above means that if someone calls of_irq_get() (or
>>>>>> platform_get_irq()), before request_irq(), then this will call
>>>>>> irq_create_of_mapping() and hence, call irq_set_irq_type. So should
>>>>>> irq_create_fwspec_mapping() be setting the type in the first place? I
>>>>>> can see it is convenient to do it here.
>>>>>
>>>>> In general there is another option - save OF-flags and pass them to
>>>>> __setup_irq() where they can be processed.
>>>>
>>>> Right, we could look at doing something like this.
>>>>
>>>>>>>> or
>>>>> [b]
>>>>>>>> 	irq_set_irq_type(irq, IRQ_TYPE_LEVEL_HIGH);
>>>>>>>> 	irq_set_chained_handler(irq, mx31ads_expio_irq_handler);
>>>>>
>>>>> option: add "flag" parameter to irq_set_chained_handler
>>>>>
>>>>>>>>
>>>>>>>> or
>>>>> [c]
>>>>>>>> 	irq_set_irq_type(alarm_irq, IRQ_TYPE_EDGE_BOTH);
>>>>>>>> 	err = devm_request_irq(&pdev->dev, alarm_irq, fan_alarm_irq_handler,
>>>>>>>> (there are ~200 occurrences of irq set_irq_type in Kernel)
>>>>>>>>
>>>>>>>> 2) if i'm not wrong, the same is valid for irq_set_irq_wake() and irq_set_affinity()
>>>>>>>>
>>>>>>>> I'm not saying all these code is correct, but that what's now in kernel :(
>>>>>>>> I've tried to test Soren's patch with omap-gpio and immediately hit case [a] :.(
>>>>>>>
>>>>>>> All functions for which are part of the public API and for which it is legal
>>>>>>> to call them without calling request_irq() (or similar) first will need to
>>>>>>> have pm_get()/pm_put().
>>>>>>
>>>>>> Right. May be we can look at the various entry points to the chip
>>>>>> operators to get a feel for which public APIs need to be handled.
>>>>>
>>>>>
>>>>> Seems yes. But we need to be very careful with this, some of functions could be
>>>>> called recursively (nested), like:
>>>>> [d]
>>>>> static int pcf857x_irq_set_wake(struct irq_data *data, unsigned int on)
>>>>> {
>>>>> 	...
>>>>> 	error = irq_set_irq_wake(gpio->irq_parent, on);
>>>>>
>>>>>
>>>>> Personally, I have nothing against irq_pm_(get|put) :) and thought about similar things
>>>>> when tried to solve the same problem for omap-gpio driver.
>>>>> But :(, I have to fall back to irq_bus_lock/sync_unlock, because of [a,b,c] - all above
>>>>> APIs surrounded by chip_bus_lock/chip_bus_sync_unlock. ([d] - I've not hit it just because
>>>>> I was lucky).
>>>>
>>>> I had a quick peek at the omap-gpio driver and I see that internally you
>>>> are using the gpio ref-count to manage RPM and use the bus-lock hooks to
>>>> invoke RPM.
>>>>
>>>> This can definitely be complex when considering all the potential paths,
>>>> but I think that we need to a look at this from a chip-ops perspective
>>>> because only the chip knows if it is accessible or not. That said, what
>>>> we need to assess is:
>>>>
>>>> 1. Which chip-ops should ONLY be called after an IRQ has been allocated
>>>>    (eg, enable/disable, mask/unmask, type, etc). These chip-ops should
>>>>    not try to control the chip PM, but should possibly WARN if called
>>>>    when  the chip is not accessible.
>>>> 2. For chip-ops that may be called without allocating an IRQ (eg.
>>>>    bus_lock, irq_suspend, etc), can these be called from an atomic
>>>>    context? If they might be called from an atomic context then these
>>>>    are the chip-ops which will cause problems as we cannot guarantee
>>>>    that all IRQ chips can support irq-safe RPM.
>>>
>>> They can't. chip_bus_lock() can sleep, so anything that locks the bus can't
>>> be called from atomic context.
>>
>> Sorry, what can't? Yes I understand that we cannot call anything that
>> locks the bus from an atomic context.
> 
> They can't be called from atomic context. The chip_bus_lock() function may
> sleep and you can't access the device without previously locking the bus.
> Since the device only needs to be powered up when it is accessed its safe to
> assume that the places where you need pm_get()/pm_put() are only called from
> non-atomic context.

Right, absolutely.

>>> One easy way out might be to always call pm_get/pm_but from
>>> bus_lock,/bus_unlock. This way the chip is guaranteed to be powered up when
>>> accessed happens. In addition pm_get is called when the IRQ is request and
>>> pm_put is called when the IRQ is release, this is to ensure the chip stays
>>> powered when it is actively monitoring the IRQ lines.
>>
>> Yes I had thought about that, but it is not quite that easy, because in
>> the case of request_irq() you don't want to pm_put() after the
>> bus_unlock(). However, the bus_lock/unlock() are good indicators of
>> different paths.
> 
> You'd call pm_get() twice in request_irq() once from bus_lock() and once
> independently, that way you still have a reference even after the bus_unlock().

Yes that is a possibility. However, there are places such as
show_interrupts() (kernel/irq/proc.c) and irq_gc_suspend() that do not
call bus_lock/unlock() which would need to be handled for PM. May be
these should also call bus_lock() as well?

Jon
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/