Date:   Tue, 21 Sep 2021 10:11:05 +0200
From:   Miquel Raynal <miquel.raynal@...tlin.com>
To:     Jonathan Cameron <jic23@...nel.org>
Cc:     Lars-Peter Clausen <lars@...afoo.de>, linux-iio@...r.kernel.org,
        linux-kernel@...r.kernel.org,
        Thomas Petazzoni <thomas.petazzoni@...tlin.com>,
        Nuno Sa <Nuno.Sa@...log.com>
Subject: Re: [PATCH v2 15/16] iio: adc: max1027: Add support for external
 triggers

Hi Jonathan,

jic23@...nel.org wrote on Mon, 20 Sep 2021 18:43:38 +0100:

> On Mon, 20 Sep 2021 10:37:39 +0200
> Miquel Raynal <miquel.raynal@...tlin.com> wrote:
> 
> > Hi Jonathan,
> > 
> > jic23@...nel.org wrote on Sat, 18 Sep 2021 18:13:08 +0100:
> >   
> > > On Wed, 15 Sep 2021 12:18:32 +0200
> > > Miquel Raynal <miquel.raynal@...tlin.com> wrote:
> > >     
> > > > Hi Jonathan, Nuno,
> > > > 
> > > > jic23@...nel.org wrote on Sun, 5 Sep 2021 17:10:46 +0100:
> > > >       
> > > > > On Thu,  2 Sep 2021 23:14:36 +0200
> > > > > Miquel Raynal <miquel.raynal@...tlin.com> wrote:
> > > > >         
> > > > > > So far the driver only supported to use the hardware cnvst trigger. This
> > > > > > was purely a software limitation.
> > > > > > 
> > > > > > The IRQ handler is already registered as being a poll function and thus
> > > > > > can be called upon external triggering. In this case, a new conversion
> > > > > > must be started, and one must wait for the data to be ready before
> > > > > > reading the samples.
> > > > > > 
> > > > > > As the same handler can be called from different places, we check the
> > > > > > value of the current IRQ with the value of the registered device
> > > > > > IRQ. Indeed, the first step is to get called with a different IRQ number
> > > > > > than ours, this is the "pullfunc" version which requests a new          
> > > > > 
> > > > > pullfunc?
> > > > >         
> > > > > > conversion. During the execution of the handler, we will wait for the
> > > > > > EOC interrupt to happen. This interrupt is handled by the same
> > > > > > helper. This time the IRQ number is the one we registered, we can in
> > > > > > this case call complete() to unlock the primary handler and return. The
> > > > > > primary handler continues executing by retrieving the data normally and
> > > > > > finally returns.          
> > > > > 
> > > > > Interesting to use the irq number..
> > > > > 
> > > > > I'm a little nervous about how this has ended up structured.
> > > > > I'm not 100% sure my understanding of how you've done it is correct.
> > > > > 
> > > > > We should have the following situation:
> > > > > 
> > > > > IRQ IN
> > > > >   |
> > > > >   v
> > > > > Trigger IRQ / EOC IRQ  (this is the spi->irq)  (currently iio_trigger_generic_data_poll_ready)
> > > > >   |              |
> > > > >   ---------      v
> > > > >   |        |   complete
> > > > >   v        v
> > > > > TrigH1    (TrigH2)   (these are the IRQs below the irq_chip IIO uses to demux triggers)
> > > > > 
> > > > > 
> > > > > So when using it's own trigger we are using an internal interrupt
> > > > > tree burried inside the IIO core.  When using it only as an EOC interrupt we shouldn't
> > > > > be anywhere near that internal interrupt chip.
> > > > > 
> > > > > So I'm surprised the IRQ matches with the spi->irq as 
> > > > > those trigH1 and trigH2 will have their own IRQ numbers.
> > > > > 
> > > > > For reference I think your architecture is currently
> > > > > 
> > > > > IRQ IN
> > > > >   |
> > > > >   v
> > > > >   Trigger IRQ
> > > > >   |
> > > > >   v
> > > > >  TRIG H1
> > > > >  Either fills the buffer or does the completion.
> > > > > 
> > > > > I am a little confused how this works with an external trigger because the Trig H1 interrupt
> > > > > should be disabled unless we are using the trigger.  That control isn't exposed to the
> > > > > driver at all.
> > > > > 
> > > > > Is my understanding right or have I gotten confused somewhere?        
> > > > 
> > > > I think the confusion comes from the fact that in the
> > > > current implementation, Trigger IRQ and EOC IRQ handlers are the same.
> > > > This comes from a possible misunderstanding in the previous review,
> > > > where I understood that you and Nuno wanted to keep using
> > > > iio_trigger_generic_data_rdy_poll() hand have a single handler in the
> > > > driver (which I think is far from optimal). I can try to split that
> > > > handler again to have two distinct paths.      
> > > That is the right thing to do.  The split should be done a little differently
> > > than you have it in v3. I've added comments to that patch.
> > > 
> > > Data ready triggers are always a little messy because we end up with a split that
> > > is:
> > > 
> > > Trigger side -  Interrupt comes in here...
> > > 
> > > --------- GENERIC IIO HANDLING ----- Take the trigger and routes it to the device code --- 
> > > 
> > > Device side - We do the data reading here.
> > > 
> > > The reason for this is that we may well have other devices using the same trigger
> > > and we want to keep the model looking the same for all devices.
> > > 
> > > A push into an iio buffer should always be on the device side of that boundary.    
> > 
> > This is much clearer, I think I have got the main idea.
> > 
> > However I have a question that is specific to the current situation. In
> > the case of this particular device, I don't really understand how
> > another device could use the same trigger than the hardware one,
> > because we have no indication of the trigger being latched. When we get
> > the information the data is already in the FIFO, this means we get the
> > information much later than when the hardware transitioned to indicate
> > a conversion request. Is it that in your model, we should be able to
> > use the EOC IRQ handler to trigger another IIO device, even though
> > this implies an additional delay?  
> 
> It's not ideal, but sometimes it is better to have somewhat consistent
> 'synchronization' between multiple devices.  You are correct that anything
> else using a data ready trigger will be a bit late - but the frequencies
> will at least be matched.  Not great but the best possible under these
> circumstances.
> 
> If it's possible to use a truely shared hardware trigger that is obviously
> better than you can do here.

Ok. This was definitely a part of the puzzle that I missed in the first
place, making harder the understanding (and the interest) of the driver
vs. core split.

Cheers,
Miquèl

Powered by blists - more mailing lists