Re: [PATCH v3 1/6] dt-bindings: iio: introduce trigger providers, consumers

[Jonathan Cameron]

On 17/03/17 15:59, Fabrice Gasnier wrote:
> On 03/15/2017 08:25 PM, Rob Herring wrote:
>> On Sun, Mar 05, 2017 at 12:13:36PM +0000, Jonathan Cameron wrote:
>>> On 05/03/17 11:43, Jonathan Cameron wrote:
>>>> On 03/03/17 06:21, Rob Herring wrote:
>>>>> On Tue, Feb 28, 2017 at 05:51:14PM +0100, Fabrice Gasnier wrote:
>>>>>> Document iio provider and consumer bindings.
>>>>>>
>>>>>> Signed-off-by: Fabrice Gasnier <fabrice.gasnier@xxxxxx>
>>>>>> ---
>>>>>>  .../devicetree/bindings/iio/iio-bindings.txt       | 38 ++++++++++++++++++++++
>>>>>>  1 file changed, 38 insertions(+)
>>>>>>
>>>>>> diff --git a/Documentation/devicetree/bindings/iio/iio-bindings.txt b/Documentation/devicetree/bindings/iio/iio-bindings.txt
>>>>>> index 68d6f8c..01765e9 100644
>>>>>> --- a/Documentation/devicetree/bindings/iio/iio-bindings.txt
>>>>>> +++ b/Documentation/devicetree/bindings/iio/iio-bindings.txt
>>>>>> @@ -95,3 +95,41 @@ vdd channel is connected to output 0 of the &ref device.
>>>>>>  		io-channels = <&adc 10>, <&adc 11>;
>>>>>>  		io-channel-names = "adc1", "adc2";
>>>>>>  	};
>>>>>> +
>>>>>> +==IIO trigger providers==
>>>>>> +Sources of IIO triggers can be represented by any node in the device
>>>>>> +tree. Those nodes are designated as IIO trigger providers. IIO trigger
>>>>>> +consumer uses a phandle and an IIO trigger specifier to connect to an
>>>>>> +IIO trigger provider.
>>>>>> +An IIO trigger specifier is an array of one or more cells identifying
>>>>>> +the IIO trigger output on a device. The length of an IIO trigger
>>>>>> +specifier is defined by the value of a #io-trigger-cells property in
>>>>>> +the IIO trigger provider node.
>>>>>> +
>>>>>> +Required properties:
>>>>>> +#io-trigger-cells:
>>>>>> +		Number of cells in an IIO trigger specifier; Typically
>>>>>> +		0 for nodes with a simple IIO trigger output.
>>>>>> +
>>>>>> +Example:
>>>>>> +	trig0: interrupt-trigger0 {
>>>>>> +		#io-trigger-cells = <0>;
>>>>>> +		compatible = "interrupt-trigger";
>>>>>> +		interrupts = <11 0>;
>>>>>> +		interrupt-parent = <&gpioa>;
>>>>>> +	}
>>>>>> +
>>>>>> +==IIO trigger consumers==
>>>>>> +Required properties:
>>>>>> +- io-triggers:	List of phandle representing the IIO trigger specifier.
>>>>>> +
>>>>>> +Optional properties:
>>>>>> +- io-trigger-names :
>>>>>> +		List of IIO trigger name strings that matches elements
>>>>>> +		in 'io-triggers' list property.
>>>>>> +
>>>>>> +Example:
>>>>>> +	some_trigger_consumer {
>>>>>> +		io-triggers = <&trig0>;
>>>>>> +		io-trigger-names = "mytrig";
>>>>>> +	}
>>>>>
>>>>> I have some reservations about this. We could just as easily add the 
>>>>> interrupt directly to the consumer node and use "trigger" for a standard 
>>>>> interrupt name. So the question is whether this extra level of 
>>>>> indirection is needed? 
>>>>
>>>> First thing to note here, is that Fabrice's use of the generic interrupt
>>>> trigger is an extremely 'unusual' one! Normal use case is that we have
> Hi Rob, Jonathan,
> 
> Yes, I agree this is unusual.
> 
>>>> a random gpio pin providing interrupts to driver triggering on random
>>>> devices - there need be no association between the two whatsoever.
>>>> So what we are doing here is 'allowing' an interrupt to provide a trigger.
>>>> It's not necessarily the one going to be used by any particular device
>>>> driver.  The decision of which trigger to use is definitely one for
>>>> userspace, not something that should be configured in to the device tree.
>>>>
>>>> For this particular case you could in theory just do it by using an interrupt
>>>> as you describe.  Ultimately though we should be able to play more complex
>>>> games with this device and having it able to handle any trigger - which 
>>>> includes those not using the direct hardware route.  It'll be up to the
>>>> driver to figure out when it can use the fast method and when it can't.
> Agreed. Still, to benefit from hw capabilities, driver needs to have a
> way to identify a particular trigger as a direct hardware route, or not
> (and then default software handling, btw, that still needs to be
> addressed in stm32-adc). DT Provider/consumer may help to achieve this.
A simple additional callback that has access to both provider and consumer
should have sufficient data to figure this out somehow.
> 
>>>>
>>>> Conversely, even when we are using this hardware route to drive the
>>>> triggering it should be possible to hang off a device to be triggered
>>>> by the interrupt via the kernel rather than directly. 
>>>>
>>>> So from a device tree point of view we are just describing the fact that
>>>> there is a pin, which may be used to trigger something.  What that something
>>>> is, is a question for userspace not the device tree.
>>>>
>>> Ah, I'm half asleep this morning.  Clearly there is a more general follow
>>> up question.  If we are arguing these are generic, why are we setting
>>> up the mapping in device tree?
>>>
>>> My gut feeling is we shouldn't be.  So I think we need the first chunk
>>> above but the latter part should be a job for userspace not the devicetree.
>>
>> So you mean keep the provider side, but get rid of the consumer? That 
>> makes sense to me.
> In case getting rid of the consumer part, I still need one way, on
> consumer side (stm32-adc) to specifically map EXTI signal in ADC
> hardware, do some checks on trigger to validate hardware route.
> I'm not sure how to handle this if I get rid of consumer part.
> Shall I use something else not mentioned here? (put trigger dt node as
> child node of stm32 adc, then use dev bus? ...)
Something along those lines might work.
> 
> Another way is as suggested by Rob in earlier comment: directly use this
> interrupt in consumer dt node, e.g. in stm32-adc node. And register
> relevant trigger and so on, from stm32-adc driver.
> Please advise.
Also a possible option, be it a little restrictive given in theory at least
we could have a situation where this trigger had hardware routes to more
than one driver.

So I don't have a answer right now.  Would need to dig a little into the
actual drivers in question...  Run out of time this weekend, but might
get some time during the week.

Brief thoughts though: the interrupt trigger obviously knows the interrupt
in question.  Can we provide some access for the ADC to which interrupts
are valid and then an optional callback like validate_trigger that
is capable of discovering that there is a hardware route?  Some sort
of handle that the trigger provides would give the device scope to see
if it knows it can do 'magic' or not, return value could indicate to
the trigger whether it is going to be doing it with 'magic' or
really providing interrupts.

Jonathan

> 
> Thanks for reviewing
> Best Regards,
> Fabrice
>>
>> Rob
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-iio" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>