[Linux-stm32] [PATCH] iio: adc: stm32-adc: fix erroneous handling of spurious IRQs

Fabrice Gasnier fabrice.gasnier at foss.st.com
Tue Jan 19 12:56:55 EST 2021


On 1/18/21 12:42 PM, Ahmad Fatoum wrote:
> Hello Jonathan,
>
> On 16.01.21 18:53, Jonathan Cameron wrote:
>> On Tue, 12 Jan 2021 16:24:42 +0100
>> Ahmad Fatoum <a.fatoum at pengutronix.de> wrote:
>>
>>> 1c6c69525b40 ("genirq: Reject bogus threaded irq requests") makes sure
>>> that threaded IRQs either
>>>   - have IRQF_ONESHOT set
>>>   - don't have the default just return IRQ_WAKE_THREAD primary handler
>>>
>>> This is necessary because level-triggered interrupts need to be masked,
>>> either at device or irqchip, to avoid an interrupt storm.
>>>
>>> For spurious interrupts, the STM32 ADC driver still does this bogus
>>> request though:
>>>   - It doesn't set IRQF_ONESHOT
>>>   - Its primary handler just returns IRQ_WAKE_THREAD if the interrupt
>>>     is unexpected, i.e. !(status & enabled_mask)
>> This seems 'unusual'.  If this is a spurious interrupt we should be
>> returning IRQ_NONE and letting the spurious interrupt protection
>> stuff kick in.
>>
>> The only reason I can see that it does this is print an error message.
>> I'm not sure why we need to go into the thread to do that given
>> it's not supposed to happen. If we need that message at all, I'd
>> suggest doing it in the interrupt handler then return IRQ_NONE;
> As described, I run into the spurious IRQ case, so I think the message is
> still useful (until that's properly fixed), but yes, it should've returned
> IRQ_NONE in that case.
>
> With these changes, IRQF_ONESHOT shouldn't be necessary, but in practice
> the driver doesn't function correctly with the primary IRQ handler threaded.
>
> Olivier, Fabrice: Are you aware of this problem?


Hi Ahmad, Jonathan,

I wasn't aware of this up to now. I confirm we've the same behavior at
our end with threadirqs=1.

Olivier and I started to look at this. Indeed, the IRQF_ONESHOT makes
the issue to disappear.
I'm not sure 100% that's for the above reasons. Please let me share some
piece of logs, analysis and thoughts.


I may miss it but, the patch "genirq: Reject bogus threaded irq
requests" seems to handle the case where no HW handler is provided, but
only the threaded part?

In the stm32-adc both are provided. Also the IRQ domain in
stm32-adc-core maybe a key here ?


We did some testing, ftrace and observed following behavior for one
capture (a single cat in_voltage..._raw) :

in stm32-adc-core, as IRQ source is still active until the IRQ thread
can execute:
- stm32_adc_irq_handler <-- generic_handle_irq
- stm32_adc_irq_handler <-- generic_handle_irq
- stm32_adc_irq_handler <-- generic_handle_irq
...

- sched_switch to the 1st IRQ thread
- stm32_adc_irq_handler <-- generic_handle_irq (again until DR get read)

- stm32_adc_isr <-- irq_forced_thread_fn (from stm32-adc)
  DR read, clears the active flag
- stm32_adc_isr <-- irq_forced_thread_fn
  wakes the 2nd IRQ thread to print an error (unexpected...)

sched_switch to the 2nd IRQ thread that prints the message.

- stm32_adc_threaded_isr <-- irq_thread_fn


So my understanding is: the cause seems to be the concurrency between

- stm32_adc_irq_handler() storm calls in stm32-adc-core
- stm32_adc_isr() call to clear the cause (forced into a thread with
threadirqs=1).

To properly work, the stm32_adc_irq_handler() should be masked in between.

As you explain, this works in this case: the call to stm32_adc_isr (in
stm32-adc) isn't longer forced threaded with IRQF_ONESHOT.

It looks like IRQF_NO_THREAD for forced threading would have similar
effect? Maybe the same would be applicable here ? (I haven't tested...)


Hopefully this helps and is similar to what you observed.

Thanks and best regards,
Fabrice

>
> Cheers,
> Ahmad
>
>>>   - stm32mp151.dtsi describes the ADC interrupt as level-triggered
>>>
>>> Fix this by setting IRQF_ONESHOT to have the irqchip mask the IRQ
>>> until the IRQ thread has finished.
>>>
>>> IRQF_ONESHOT also has the effect that the primary handler is no longer
>>> forced into a thread. This makes the issue with spurious interrupts
>>> interrupts disappear when reading the ADC on a threadirqs=1 kernel.
>>> This used to result in following kernel error message:
>>>
>>> 	iio iio:device1: Unexpected IRQ: IER=0x00000000, ISR=0x0000100e
>>> or
>>> 	iio iio:device1: Unexpected IRQ: IER=0x00000004, ISR=0x0000100a
>>>
>>> But with this patch applied (or threaded IRQs disabled), this no longer
>>> occurs.
>>>
>>> Cc: Lucas Stach <l.stach at pengutronix.de>
>>> Reported-by: Holger Assmann <has at pengutronix.de>
>>> Fixes: 695e2f5c289b ("iio: adc: stm32-adc: fix a regression when using dma and irq")
>>> Signed-off-by: Ahmad Fatoum <a.fatoum at pengutronix.de>
>>> ---
>>>  drivers/iio/adc/stm32-adc.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/iio/adc/stm32-adc.c b/drivers/iio/adc/stm32-adc.c
>>> index c067c994dae2..7e0e21c79ac8 100644
>>> --- a/drivers/iio/adc/stm32-adc.c
>>> +++ b/drivers/iio/adc/stm32-adc.c
>>> @@ -1910,7 +1910,7 @@ static int stm32_adc_probe(struct platform_device *pdev)
>>>  
>>>  	ret = devm_request_threaded_irq(&pdev->dev, adc->irq, stm32_adc_isr,
>>>  					stm32_adc_threaded_isr,
>>> -					0, pdev->name, indio_dev);
>>> +					IRQF_ONESHOT, pdev->name, indio_dev);
>>>  	if (ret) {
>>>  		dev_err(&pdev->dev, "failed to request IRQ\n");
>>>  		return ret;
>>



More information about the linux-arm-kernel mailing list