Re: [PATCH] staging: iio: ad7606: use iio_device_{claim|release}_direct_mode()

[Jonathan Cameron <jic23@kernel.org>]

On 06/04/16 19:06, Lars-Peter Clausen wrote:
> On 04/03/2016 11:09 AM, Jonathan Cameron wrote:
>> On 01/04/16 17:53, Alison Schofield wrote:
>>> Two instances are moved to the new claim/release API:
>>>
>>> In the first instance, the driver was using mlock followed by
>>> iio_buffer_enabled(). Replace that code with the new API to guarantee
>>> the device stays in direct mode. There is no change in driver behavior.
>>>
>>> In the second instance, the driver was not using mlock to hold the
>>> device in direct mode, but should have been.  Here we introduce the
>>> new API to guarantee direct mode. This is a change in driver behavior.
>>>
>>> Signed-off-by: Alison Schofield <amsfield22@gmail.com>
>>> ---
>>>  drivers/staging/iio/adc/ad7606_core.c | 15 ++++++++-------
>>>  1 file changed, 8 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/drivers/staging/iio/adc/ad7606_core.c
> b/drivers/staging/iio/adc/ad7606_core.c
>>> index 6dbc811..f914b8d 100644
>>> --- a/drivers/staging/iio/adc/ad7606_core.c
>>> +++ b/drivers/staging/iio/adc/ad7606_core.c
>>> @@ -88,12 +88,12 @@ static int ad7606_read_raw(struct iio_dev *indio_dev,
>>>
>>>  	switch (m) {
>>>  	case IIO_CHAN_INFO_RAW:
>>> -		mutex_lock(&indio_dev->mlock);
>>> -		if (iio_buffer_enabled(indio_dev))
>>> -			ret = -EBUSY;
>>> -		else
>>> -			ret = ad7606_scan_direct(indio_dev, chan->address);
>>> -		mutex_unlock(&indio_dev->mlock);
>>> +		ret = iio_device_claim_direct_mode(indio_dev);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		ret = ad7606_scan_direct(indio_dev, chan->address);
>>> +		iio_device_release_direct_mode(indio_dev);
>>>
>>>  		if (ret < 0)
>>>  			return ret;
>>> @@ -411,8 +411,9 @@ static irqreturn_t ad7606_interrupt(int irq, void
> *dev_id)
>>>  	struct iio_dev *indio_dev = dev_id;
>>>  	struct ad7606_state *st = iio_priv(indio_dev);
>>>
>>> -	if (iio_buffer_enabled(indio_dev)) {
>>> +	if (!iio_device_claim_direct_mode(indio_dev))  {
>>>  		schedule_work(&st->poll_work);
>>> +		iio_device_release_direct_mode(indio_dev);
>> Unfortunately this won't work.  That interrupt is still in traditional non
>> threaded form.  This will take a mutex in a top half interrupt handler
>> where a sleep cannot occur.
>>
>> I'm just wondering how expensive it would be to fix this by moving that over
>> to a threaded handler.  In the poll_work case (buffer) it would be cleaner
> to do
>> so. I'm really confused what the intended interrupt handler
>> is in here.  I 'think' the sequence is:
>>
>> Trigger fires the convst pin whether in top half or the bottom half of
>> a threaded interrupt, but not both - I guess this works, if it is rather
>> 'unusual'.
>>
>> We then get a interrupt to indicate that it has finished conversion and that
>> filters through to actually fill the buffer via a traditional top half /
>> bottom half interrupt handler.
>>
>> So if we were to convert that to a threaded interrupt (with no top half / non
>> threaded part), we could drop the schedule_work and just call
>> ad7606_poll_bh_to_ring from the thread handler.
>>
>> In the direct read case I doubt we care about the delay in dropping to a
>> thread prior to signalling the data is ready.
>>
>> Can't think why this driver is still in staging :)
> 
> Yeah, we should leave this driver out from the conversion for now. The whole
> convst pin handling need a major rework. It shouldn't really be in the
> driver and usually you wouldn't want to use to use a GPIO and software timer
> since that gives you way to much jitter for good results. You'd probably use
> something like a PWM.
Agreed, you would normally move this into hardware, but we'd still need to allow
for the possibility of it being software triggered - even if ugly.

Anyhow, lets park this for now.

J
> 
>>
>> Lars, any interest from Analog in getting this one cleaned up?  Also
>> do you have any test hardware, if we mess around with this interrupt handling?
> 
> I have the hardware somewhere in some storage bay, but just converting this
> over to threaded interrupt handling is not really a solution. So, if you
> want to get rid of the iio_buffer_enabled() in the interrupt handler a
> simple solution is to register preenable and postdisable callbacks where you
> set a flag in the driver struct to indicate that it is in buffered mode or not.
>