Message-ID: <20210918183920.236f4f0d@jic23-huawei>
Date:   Sat, 18 Sep 2021 18:39:20 +0100
From:   Jonathan Cameron <jic23@...nel.org>
To:     Miquel Raynal <miquel.raynal@...tlin.com>
Cc:     "Sa, Nuno" <Nuno.Sa@...log.com>,
        Lars-Peter Clausen <lars@...afoo.de>,
        "linux-iio@...r.kernel.org" <linux-iio@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Thomas Petazzoni <thomas.petazzoni@...tlin.com>
Subject: Re: [PATCH v2 15/16] iio: adc: max1027: Add support for external
 triggers

On Wed, 15 Sep 2021 17:45:07 +0200
Miquel Raynal <miquel.raynal@...tlin.com> wrote:

> Hi Nuno, Jonathan,
> 
> Nuno.Sa@...log.com wrote on Mon, 6 Sep 2021 09:53:15 +0000:
> 
> > > -----Original Message-----
> > > From: Jonathan Cameron <jic23@...nel.org>
> > > Sent: Sunday, September 5, 2021 6:11 PM
> > > To: Miquel Raynal <miquel.raynal@...tlin.com>
> > > Cc: Lars-Peter Clausen <lars@...afoo.de>; linux-iio@...r.kernel.org;
> > > linux-kernel@...r.kernel.org; Thomas Petazzoni
> > > <thomas.petazzoni@...tlin.com>; Sa, Nuno <Nuno.Sa@...log.com>
> > > Subject: Re: [PATCH v2 15/16] iio: adc: max1027: Add support for
> > > external triggers
> > > 
> > > [External]
> > > 
> > > 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 also can't see a path in which the eoc interrupt will get fired for
> > > raw_reads.
> > > 
> > > Could you talk me through how that works currently?  
> 
> I forgot to do this, I think you misunderstood the following patch:
> "iio: adc: max1027: Don't just sleep when the EOC interrupt is
> available"
> 
> As I am having deep troubles reworking this once again, I will try to
> explain how this is build below and look for your feedback on it.

Unfortunately I'm not successfully managing to convey what I am trying
to get you to change this to...

> 
> > > 
> > > I suspect part of the confusion here is that this driver happens to be
> > > using the
> > > standard core handler iio_trigger_generic_data_rdy_poll which hides
> > > away that
> > > there are two interrupt handlers in a normal IIO driver for a device with
> > > a
> > > trigger and buffered mode.
> > > 1 for the trigger and 1 for the buffer.  Whether the buffer one is a
> > > result
> > > of the trigger one (via iio_poll_trigger) is down to whether the device
> > > is
> > > using it's own trigger or not.
> > >   
> 
> [...]
>  
> > I'm also confused by this. Going through the series, I was actually
> > thinking that raw_reads were in fact using the EOC IRQ until I realized
> > that 'complete()' was being called from the trigger handler... So,
> > I'm also not sure how is this supposed to work?  
> 
> I renamed this handler with a generic name, because it actually serves
> several different purposes, see below.
> 
> > But I'm probably
> > missing something as I guess you tested this and how I'm understanding
> > things, you should have gotten timeouts for raw_reads.
> > 
> > Anyways, as Jonathan said, I was also expecting to see the 'complete()' call
> > from the device IRQ handler. Other thing than that is just asking for trouble
> > :). 
> > 
> > - Nuno Sá  
> 
I'm going to take a multi pronged approach to trying to get us on the same
page design wise.   I've replied to v3 with code snippets that will hopefully
convey the explanation I'm giving here.

Data ready triggers always provide some blurring of the nice clean
trigger -> IIO CORE HANDLING -> device handling to grab data and push it out..
but we have a fairly standard model for this and whilst limited in functionality
I think the max1027 was keeping to this model.

> So here is how I understand the device:
> * During a raw read:
>   The IRQ indicates an EOC.
> * During a triggered read (internal trigger):
>   The driver has no knowledge of the trigger being fired, it only
>   sees the IRQ firing at EOC. This means the internal trigger cannot be
>   used without the IRQ.
> 
> Now here is what I understand from the IIO subsystem and what I tried
> to implement:
> * Perform a raw read:
>   The driver needs to setup the channels, start the operation, wait for
>   the end of the conversion, return the value. This is all done in the
>   ->raw_read() hook in process context. Raw reads can either use the  
>   EOC IRQ if wired or just wait for a sufficiently large amount of
>   time.
> * Perform a triggered read (internal trigger):
>   The device will get triggered by a hardware event that is out of
>   control from the driver. The driver is only aware of the IRQ firing
>   when the conversion is done. It then must push the samples to a set
>   of buffers and notify the IIO core.

True, but there are two steps to this.
1) The trigger IRQ fires.  That then calls a tree of other interrupts
(iio_trigger_poll()).
One of that tree of interrupts is attached by the driver.
That interrupt handler is then called to actually take the data indicated
by the EOC interrupt and push it to the buffer.  Note this second interrupt
is registered as part of iio_triggered_buffer_setup() which is why that
function takes interrupt handlers.

> * Perform a triggered read (external trigger):
>   This looks very much like a raw read from the driver point of view,
>   the difference being that, when the external trigger fires, the IIO
>   core will first call iio_pollfunc_store_time() as hard IRQ handler
>   and then calls a threaded handler that is supposed to configure the
>   channels, start the conversions, wait for it and again push the
>   samples when their are ready. These two handlers are registered by
>   devm_iio_triggered_buffer_setup().


That is true in your code, but it is not the correct way to do this because
in the event of a trigger other than the EOC one, we should not have the
interrupt to indicate the end of the read we are manually starting calling
into the handler for the trigger (as it wasn't the trigger!) - it should
be handled at the level above that.

If you have a different trigger than the EOC one provided by this driver then
the top level interrupt handler (the one that actually handles the EOC interrupt
should not call iio_trigger_generic_data_ready_poll() because we need to do
something different in that interrupt handler.  It is this handler that
you should modify so that it does different things depending on whether or
not the EOC trigger is in use. (you can check this using iio_trigger_using_own(iio_dev))

I've tried to find a simple example of this, but they are all rather convoluted for
various device specific reasons (getting the best possible timestamp for example)

https://elixir.bootlin.com/linux/latest/source/drivers/iio/pressure/zpa2326.c#L786
more or less follows the model I'm describing with some rather fiddly timestamp
handling..

> 
> There is an additional level of complexity as, my hardware does not
> feature this EOC IRQ (it is not wired and thus cannot be used). I want
> to be able to also support the absence of IRQ which is not necessary
> for any operation but the internal trigger use.
> 
> How I ended-up implementing things in v2:
> * Raw read:
>   Start the conversion.
>   Wait for the conversion to be done:
>   - With IRQ: use wait_for_completion(), the IRQ will fire, calling
>     iio_trigger_generic_data_rdy_poll() [1], calling complete() in this
>     situation.
>   - Without IRQ: busy wait loop.
> 
> [1] Maybe this is the thing that is bothering you, using the internal
> IIO trigger interrupt tree logic to handle a regular EOC situation.

Yes.

> In
> all drivers that I had a chance to look at, it's always done like that
> but perhaps it was bad luck and the more I look at it, the less I
> understand its use. I will propose an alternative where the hard IRQ
> handler really is the EOC handler and eventually calls a threaded
> handler in the case of an internal triggering to push the samples.
> I don't see the point of using iio_trigger_generic_data_rdy_poll()
> anymore (maybe I'm wrong, tell me after reading v3?).



> 
> I ended-up writing this driver this way because I thought that I had to
> provide a single IRQ handler and use
> iio_trigger_generic_data_rdy_poll() but now I think I either
> misinterpreted the comments or it was bad advise. I fell that the
> driver is much simpler to understand in v3 where there is:
> - one hard IRQ handler registered as the primary interrupt handler:
>   its purpose is to handle EOC conditions
> - one threaded handler registered as the secondary interrupt handler:
>   it will only be triggered when using the internal trigger, its purpose
>   is to read and push the samples
> - one threaded handler registered as the external trigger handler:
>   it will start the conversion, wait for EOC (either by busy waiting if
>   there is no IRQ or by waiting for completion otherwise - complete()
>   being called in the primary IRQ handler), read the data and push it
>   to the buffers.
This nearly works, but I think the trigger reference counting will be broken
because you call iio_trigger_notify_done() without calling iio_trigger_poll().
You could cheat and drop the notify_done() - which is the solution used when
we have fifos involved and hence have to do a still more complex dance but
that isn't the right solution here.

This is more complex than should be needed and doesn't fit the model IIO
uses to separate triggers and data capture.

A few 'rules' which get bent occasionally, particularly when a driver only
supports it's own trigger.
* iio_trigger_poll() only called in an interrupt handler that is handling
  the interrupt which is the trigger in use.  So in this case, only when
  the driver is configured to use the EOC based trigger.
* iio_push_to_buffers_*(), iio_trigger_notify_done() only called from a
  trigger handler which is registered with iio_triggered_buffer_setup))

To keep to this set of rules we need two paths in the trigger irq handler
and in the trigger handler (the one on the 'device' side of the internal
irq chip that IIO uses to 'broadcast' triggers).

I may well have messed up details in the following but hopefully I manage
to convey the basic approach I'm suggesting.

So replace that iio_trigger_generic_data_rdy_poll() with one that does
something along the lines of

irqreturn_t irq_handler(int irq, void *private)
{
	struct iio_dev *indio_dev = private;
	struct max1027_state *st = iio_priv(indio_dev);

	if (!iio_buffer_enabled(indio_dev) ||
	    !iio_trigger_using_own(trigger) {
		complete(st->completion);
		return IRQ_HANDLED;
	}
	/* So we only use this as a trigger - if it is the trigger! */
	return iio_trigger_generic_data_rdy_poll(st->trig, irq); 	
}

registered in the 
dev_request_threaded_irq() call with the private parameter changed to indio_dev
as you have done in v3.

and

irqreturn_t trigger_handler(void *private, int irq)
{
	struct iio_poll_func *pf = private;
	struct io_dev *indio_dev = pf->indio_dev;

	if (!iio_trigger_using_own(indio_dev)) {
		ret = max1027_configure_chans_and_start(indio_dev);
		if (ret)
			return IRQ_HANDLED;

		ret = max1027_wait_eoc(indio_dev);
		if (ret)
			return IRQ_HANDLED;

	 	ret = max1027_read_scan(indio_dev);
		if (ret)
			return IRQ_HANDLED;	
		iio_trigger_notify_done(indio_dev);
		return IRQ_HANDLED;
	} else {
		ret = max1027_read_scan(indio_dev);
		if (ret)
			dev_err(&indio_dev->dev,
			"Cannot read scanned values (%d)\n", ret);
	
		iio_trigger_notify_done(indio_dev->trig);
	
		return IRQ_HANDLED;
	}
/* Feel free to rearrange as you like! */

}

> 
> If this implementation does not fit your requirements, I will really
> need more advanced advises about what you require, what I should avoid
> and perhaps what is wrong in the above explanation :)
> 
> Thanks,
> Miquèl

Powered by blists - more mailing lists