RE: [PATCH 14/16] iio: adc: max1027: Consolidate the end of conversion helper

["Sa, Nuno" <Nuno.Sa@analog.com>]

Hi Miquel,

> From: Miquel Raynal <miquel.raynal@bootlin.com>
> Sent: Thursday, September 2, 2021 5:13 PM
> To: Sa, Nuno <Nuno.Sa@analog.com>
> Cc: Jonathan Cameron <jic23@kernel.org>; Lars-Peter Clausen
> <lars@metafoo.de>; Thomas Petazzoni
> <thomas.petazzoni@bootlin.com>; linux-iio@vger.kernel.org; linux-
> kernel@vger.kernel.org
> Subject: Re: [PATCH 14/16] iio: adc: max1027: Consolidate the end of
> conversion helper
> 
> Hi Nuno,
> 
> "Sa, Nuno" <Nuno.Sa@analog.com> wrote on Mon, 30 Aug 2021
> 12:44:48
> +0000:
> 
> > > -----Original Message-----
> > > From: Jonathan Cameron <jic23@kernel.org>
> > > Sent: Monday, August 30, 2021 12:37 PM
> > > To: Miquel Raynal <miquel.raynal@bootlin.com>
> > > Cc: Lars-Peter Clausen <lars@metafoo.de>; Thomas Petazzoni
> > > <thomas.petazzoni@bootlin.com>; linux-iio@vger.kernel.org;
> linux-
> > > kernel@vger.kernel.org
> > > Subject: Re: [PATCH 14/16] iio: adc: max1027: Consolidate the end
> of
> > > conversion helper
> > >
> > > On Wed, 18 Aug 2021 13:11:37 +0200
> > > Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> > >
> > > > Now that we have a dedicated handler for End Of Conversion
> > > interrupts,
> > > > let's create a second path:
> > > > - Situation 1: we are using the external hardware trigger, a
> > > conversion
> > > >   has been triggered and the ADC pushed the data to its FIFO, we
> > > need to
> > > >   retrieve the data and push it to the IIO buffers.
> > > > - Situation 2: we are not using the external hardware trigger,
> hence
> > > we
> > > >   are likely waiting in a blocked thread waiting for this interrupt to
> > > >   happen: in this case we just wake up the waiting thread.
> > > >
> > > > Signed-off-by: Miquel Raynal <miquel.raynal@bootlin.com>
> > > > ---
> > > >  drivers/iio/adc/max1027.c | 20 +++++++++++++++++---
> > > >  1 file changed, 17 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/drivers/iio/adc/max1027.c
> b/drivers/iio/adc/max1027.c
> > > > index 8d86e77fb5db..8c5995ae59f2 100644
> > > > --- a/drivers/iio/adc/max1027.c
> > > > +++ b/drivers/iio/adc/max1027.c
> > > > @@ -235,6 +235,7 @@ struct max1027_state {
> > > >  	struct iio_trigger		*trig;
> > > >  	__be16				*buffer;
> > > >  	struct mutex			lock;
> > > > +	bool				data_rdy;
> > > >  	bool				cnvst_trigger;
> > > >  	u8				reg ____cacheline_aligned;
> > > >  };
> > > > @@ -243,12 +244,22 @@ static
> > > DECLARE_WAIT_QUEUE_HEAD(max1027_queue);
> > > >
> > > >  static int max1027_wait_eoc(struct iio_dev *indio_dev)
> > > >  {
> > > > +	struct max1027_state *st = iio_priv(indio_dev);
> > > >  	unsigned int conversion_time =
> > > MAX1027_CONVERSION_UDELAY;
> > > > +	int ret;
> > > >
> > > > -	if (indio_dev->active_scan_mask)
> > > > -		conversion_time *= hweight32(*indio_dev-
> > > >active_scan_mask);
> > > > +	if (st->spi->irq) {
> > > > +		ret =
> > > wait_event_interruptible_timeout(max1027_queue,
> > > > +						       st->data_rdy, HZ /
> > > 1000);
> > > > +		st->data_rdy = false;
> > > > +		if (ret == -ERESTARTSYS)
> > > > +			return ret;
> > > > +	} else {
> > > > +		if (indio_dev->active_scan_mask)
> > > > +			conversion_time *= hweight32(*indio_dev-
> > > >active_scan_mask);
> > > >
> > > > -	usleep_range(conversion_time, conversion_time * 2);
> > > > +		usleep_range(conversion_time, conversion_time * 2);
> > > > +	}
> > > >
> > > >  	return 0;
> > > >  }
> > > > @@ -481,6 +492,9 @@ static irqreturn_t
> > > max1027_eoc_irq_handler(int irq, void *private)
> > > >  	if (st->cnvst_trigger) {
> > > >  		ret = max1027_read_scan(indio_dev);
> > > >  		iio_trigger_notify_done(indio_dev->trig);
> > > > +	} else {
> > > > +		st->data_rdy = true;
> > > > +		wake_up(&max1027_queue);
> > >
> > > I can't see why a queue is appropriate for this.  Use a completion
> and
> > > have
> > > one per instance of the device.  No need for the flag etc in that
> case as
> > > complete() means we have had an interrupt.
> > >
> >
> > In the case that 'st-> cnvst_trigger' is not set but the spi IRQ
> > is present, we will wait until we get 'wake_up()' called from here. I
> wonder if
> > that is a good idea as the device own trigger is not being used. FWIW,
> I think this
> > sync logic is a bit confusing... I would still use the normal trigger
> infrastructure
> > ('iio_trigger_generic_data_rdy_poll()') and use the 'cnvst_trigger'
> flag in the
> > trigger handler to manually start conversions + wait till eoc. But I
> might be missing
> > something though.
> 
> I implemented it your way, but I think I found a situation that was not
> fully handled (the 3rd), which makes the handler very complicated
> as we need to handle all the following cases:
> 1/ no trigger, irq enabled -> single read EOC interrupt
> 2/ external trigger, no irq -> handle the whole conversion process
> 3/ external trigger, irq enabled -> handle the whole conversion process
>    but also have a dedicated condition to handle the EOC interrupt
>    properly (fortunately this is a threaded handler that can be
>    preempted): we need to wait from the handler itself that the
>    handler gets called again: the first time it is executed as
>    "pollfunc", the second time as "EOC interrupt". In the second
>    instance, call complete() in order to deliver the first running
>    instance of the handler and continue until the reading part.
> 4/ cnvst trigger, irq enabled -> only reads the data.
> 5/ cnvst trigger, irq disabled -> not possible.
> 
> I added a lot of comments to make it clearer.
> 
> > Regarding this handler, I just realized that this is the hard IRQ handler
> which
> > might end up calling 'max1027_read_scan()' which in turn calls
> 'spi_read()'. Am I
> > missing something here?
> 
> I renamed it to make it clear, but this is already a threaded handler.
> 

Hmm, I think I get what you're trying to do.... FWIW, I think you're just going
into a lot of trouble here for scenario 3 (I assume external trigger is something
else other the device own one). IMO, I would just assume that if we are using
an external trigger we have to wait (sleep) for the end of conversion (i.e, I would
not care about the IRQ in this case). It would make things much more simpler and
I guess it should be expected that if some user is deliberately not using the
device own trigger, will have to wait more for scans.

I cannot also see a reason why someone would want to use some
external trigger if the device one is available... Does it really make sense?

- Nuno Sá