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

From: Miquel Raynal
Date: Fri Sep 03 2021 - 10:46:55 EST


Hi Nuno,

Nuno.Sa@xxxxxxxxxx wrote on Fri, 3 Sep 2021 14:28:52 +0000:

> Hi Miquel,
>
> > From: Miquel Raynal <miquel.raynal@xxxxxxxxxxx>
> > Sent: Thursday, September 2, 2021 5:13 PM
> > To: Sa, Nuno <Nuno.Sa@xxxxxxxxxx>
> > Cc: Jonathan Cameron <jic23@xxxxxxxxxx>; Lars-Peter Clausen
> > <lars@xxxxxxxxxx>; Thomas Petazzoni
> > <thomas.petazzoni@xxxxxxxxxxx>; linux-iio@xxxxxxxxxxxxxxx; linux-
> > kernel@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH 14/16] iio: adc: max1027: Consolidate the end of
> > conversion helper
> >
> > Hi Nuno,
> >
> > "Sa, Nuno" <Nuno.Sa@xxxxxxxxxx> wrote on Mon, 30 Aug 2021
> > 12:44:48
> > +0000:
> >
> > > > -----Original Message-----
> > > > From: Jonathan Cameron <jic23@xxxxxxxxxx>
> > > > Sent: Monday, August 30, 2021 12:37 PM
> > > > To: Miquel Raynal <miquel.raynal@xxxxxxxxxxx>
> > > > Cc: Lars-Peter Clausen <lars@xxxxxxxxxx>; Thomas Petazzoni
> > > > <thomas.petazzoni@xxxxxxxxxxx>; linux-iio@xxxxxxxxxxxxxxx;
> > linux-
> > > > kernel@xxxxxxxxxxxxxxx
> > > > 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@xxxxxxxxxxx> 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@xxxxxxxxxxx>
> > > > > ---
> > > > > 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 thought about that, but this means playing with enabling/disabling
IRQs, which in the end I fear could be almost as verbose :/

Can you look at the new implementation that I proposed and give me your
feedback on it?
[PATCH v2 15/16] iio: adc: max1027: Add support for external triggers

> 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?

I would say that "genericity" is a good enough reason for that, but in
practice I agree with you.

Thanks,
Miquèl