Re: [PATCH v10 10/10] iio: accel: mma8452: Support interrupt sharing

[Esben Haabendal]

"Jonathan Cameron" <jic23@xxxxxxxxxx> writes:

> On Mon, 28 Sep 2026 12:15:57 +0200
> Esben Haabendal <esben@xxxxxxxxxx> wrote:
>
>> "Andy Shevchenko" <andriy.shevchenko@xxxxxxxxx> writes:
>>
>> > On Mon, Sep 28, 2026 at 10:26:26AM +0200, Esben Haabendal wrote:
>> >> Adding support for sharing interrupt line with other device requires the
>> >> interrupt handler to handle runtime PM suspension properly, ignoring the
>> >> irq if the device is suspended (maybe even off). And while at it, we use
>> >> the PM reference to ensure we do not get suspended while processing an irq.
>> >>
>> >> In order to prevent the chip from raising irq while suspended (that is when
>> >> using fixed regulator, where suspend just means setting the device in
>> >> STANDBY mode), we disable all interrupt sources by clearing CTRL_REG4, and
>> >> then restores the value again when resuming.
>> >>
>> >> The scoped_guard in mma8452_runtime_suspend() is changed to a plain
>> >> mutex_lock() instead, both to prevent mixing guards and goto, but also to
>> >> ensure that we stay in STANDBY mode while writing to CTRL_REG4 and all the
>> >> way up to disabling the device as much as possible.
>> >>
>> >> With that in place, it is safe to add the IRQF_SHARED flag.
>> >>
>> >> Keep in mind that the device by default is using push-pull for the irq pin,
>> >> which might require additional hardware design to allow interrupt sharing.
>> >
>> > ...
>> >
>> >> +	pm_status = pm_runtime_get_if_active(dev);
>> >> +	if (pm_status == 0)
>> >> +		/* device is powered down */
>> >> +		return IRQ_NONE;
>> >
>> >> +	if (IS_ENABLED(CONFIG_PM) && pm_status < 0)
>> >> +		/* runtime PM was disabled, possibly suspending */
>> >> +		return IRQ_HANDLED;
>> >
>> > This is very interesting part. I bet this will be the first driver using this.
>> > A big question "why?"
>>
>> Because as I understood it, I needed to add this for the patch to be
>> accepted [1]. It is definitely possible that I have misunderstood that.
>>
>> The technical reason I believe is that during system suspend,
>> pm_runtime_force_suspend() disabled runtime PM, and if a shared
>> interrupt durings this, pm_runtime_get_if_active returns -EINVAL, and we
>> could therefore end up doing I2C reads on unpowered hardware.
>>
>> Per that reasoning, we probably should be doing this in other drivers as
>> well.
>
> I suspect that the combination of a shared interrupt and runtime pm
> is moderately rare and even when it is there are drivers doing their
> own tracking to avoid this sort of corner case.
>
> There are a lot of these nasty corners in runtime pm :(
>
> This seems like a reasonable solution to me, but given
> all the discussion on this version was around this patch I'm going
> to pick up 1-9 to but ask you to send a new version of this one to
> make sure we get maximum eyes on it.  Please include the PM maintainer
> as well in the +CC in case we have this all wrong!

Okay. Anything that I should change before sending a new version? And
should I reset to version 1, or should it be v11?

/Esben

>
>>
>> >> +	/*
>> >> +	 * pm_status is now 1 or -EINVAL (with CONFIG_PM not enabled). If
>> >> +	 * pm_status==1, runtime PM is enabled and device is RPM_ACTIVE. If
>> >> +	 * pm_status==-EINVAL, runtime PM is build-time disabled (i.e. CONFIG_PM
>> >> +	 * not enabled), and we can/must assume device is active.
>> >> +	 */
>> >> +
>> >>  	src = i2c_smbus_read_byte_data(data->client, MMA8452_INT_SRC);
>> >>  	if (src < 0)
>> >> -		return IRQ_NONE;
>> >> +		goto out_runtime_put;
>> >>
>> >>  	if (!(src & (data->chip_info->enabled_events | MMA8452_INT_DRDY)))
>> >> -		return IRQ_NONE;
>> >> +		goto out_runtime_put;
>> >>
>> >>  	if (src & MMA8452_INT_DRDY) {
>> >>  		iio_trigger_poll_nested(indio_dev->trig);
>> >> @@ -1120,6 +1139,10 @@ static irqreturn_t mma8452_interrupt(int irq, void *p)
>> >>  		ret = IRQ_HANDLED;
>> >>  	}
>> >>
>> >> +out_runtime_put:
>> >> +	if (pm_status > 0)
>> >> +		pm_runtime_put_autosuspend(dev);
>> >> +
>> >>  	return ret;
>> >>  }
>> >
>> > ...
>> >
>> >> +	pm_runtime_enable(dev);
>> >> +	pm_runtime_set_autosuspend_delay(dev, MMA8452_AUTO_SUSPEND_DELAY_MS);
>> >> +	pm_runtime_use_autosuspend(dev);
>> >
>> > I don't see the respective _dont_use_autosuspend() call anywhere.
>>
>> Ah. I was not aware it was attached to a usage counter as well. I will
>> add a pm_runtime_dont_use_autosuspend() to the runtime_suspend: error handling.
>>
>> > ...
>> >
>> > I am wondering how many of the above possible scenarios you were able
>> > to test.
>>
>> Not enough. I haven't done any additional instrumentation to cause all
>> the various error conditions that is handled here in _probe().
>>
>> > ...
>> >
>> > P.S.  I bet you can now make a presentation "PM runtime in Linux and
>> > why it is so hard."
>>
>> LOL. I am getting there, yes. I have definitely learned a lot about how
>> runtime PM works. Especially since I was mostly starting from scratch on
>> that.
>>
>> But i fear it would easily end up being quite a chaotic presentation :)
>
> :)
>
> I've vaguely be wondering if we need a bunch of self test + examples
> to poke all the more crazy corners of runtime pm.  That way we can
> at least be sure that if we do finally understand it we aren't going
> to get any regressions.

Not a bad idea :)

/Esben