[Linux-stm32] [PATCH][RFC] iio: core: add a class hierarchy on iio device lock

[Jonathan Cameron]

On Tue, 22 Oct 2019 12:53:12 +0100
Jonathan Cameron <jic23 at kernel.org> wrote:

> On Tue, 15 Oct 2019 23:11:43 +0200
> Lars-Peter Clausen <lars at metafoo.de> wrote:
> 
> > On 10/14/19 5:59 PM, Olivier MOYSAN wrote:  
> > > Hello Jonathan,
> > > 
> > > Thanks for your comment.
> > > 
> > > On 10/12/19 10:57 AM, Jonathan Cameron wrote:    
> > >> On Fri, 11 Oct 2019 17:13:14 +0200
> > >> Olivier Moysan <olivier.moysan at st.com> wrote:
> > >>    
> > >>> The aim of this patch is to correct a recursive locking warning,
> > >>> detected when setting CONFIG_PROVE_LOCKING flag (as shown in message below).
> > >>> This message was initially triggered by the following call sequence
> > >>> in stm32-dfsdm-adc.c driver, when using IIO hardware consumer interface.
> > >>>
> > >>> in stm32_dfsdm_read_raw()
> > >>> 	iio_device_claim_direct_mode
> > >>> 		mutex_lock(&indio_dev->mlock);			-> lock on dfsdm device
> > >>> 	iio_hw_consumer_enable
> > >>> 		iio_update_buffers
> > >>> 			mutex_lock(&indio_dev->mlock);		-> lock on hw consumer device    
> > >> Hmm.  I'm not sure I follow the logic.  That lock is
> > >> for one thing and one thing only, preventing access
> > >> to the iio device that are unsafe when it is running
> > >> in a buffered mode.  We shouldn't be in a position where
> > >> we both say don't do this if we are in buffered mode, + enter
> > >> buffered mode whilst doing this, or we need special functions
> > >> for entering buffering mode if in this state.  We are in
> > >> some sense combining internal driver logic with overall
> > >> IIO states.  IIO shouldn't care that the device is using
> > >> the same methods under the hood for buffered and non
> > >> buffered operations.
> > >>
> > >> I can't really recall how this driver works.   Is it actually
> > >> possible to have multiple hw_consumers at the same time?
> > >>
> > >> So do we end up with multiple buffers registered and have
> > >> to demux out to the read_raw + the actual buffered path?
> > >> Given we have a bit of code saying grab one sample, I'm
> > >> going to guess we don't...
> > >>
> > >> If so, the vast majority of the buffer setup code in IIO
> > >> is irrelevant here and we just need to call a few of
> > >> the callbacks from this driver directly... (I think
> > >> though I haven't chased through every corner.
> > >>
> > >> I'd rather avoid introducing this nesting for a corner
> > >> case that makes no 'semantic' sense in IIO as it leaves us
> > >> in two separate states at the same time that the driver
> > >> is trying to make mutually exclusive.  We can't both
> > >> not be in buffered mode, and in buffered mode.
> > >>
> > >> Thanks and good luck with this nasty corner!
> > >>
> > >> Jonathan
> > >>    
> > > Here I consider the following use case:
> > > A single conversion is performed. The dfsdm (filter) is chained with a 
> > > front-end, which can be an ADC or a sensor. So we have two IIO devices, 
> > > the dfsdm and its front-end handled through the hw consumer interface.
> > > 
> > > You are right. There is something wrong here, in buffered/non-buffered 
> > > mode mixing.
> > > iio_hw_consumer_enable() call is used to enable the front-end device. 
> > > But this interface is intended for buffered mode.
> > > So this is not coherent with the expected single conversion mode, 
> > > indeed. Another interface is required to manage the front-end device. I 
> > > have a poor knowledge of iio framework, but it seems to me that there is 
> > > no interface to manage this.
> > > 
> > > My understanding regarding mlock, is that it is used to protect the 
> > > state of the iio device.
> > > I we want to do a conversion from the chained devices, I think we need 
> > > to activate the first device
> > > and keep it performing conversion, as long as the second device has done 
> > > its conversion.
> > > We need to protect both devices, and we should have to do it in a nested 
> > > way.
> > > So, I guess that anyway, nested mutexes would be required in this case.
> > >    
> > 
> > Others like regmap have solved this by having a lockclass per instance.
> > Although that is not ideal either since it will slow down lockdep.
> > 
> > See
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/regmap.h#n629  
> 
> It'll take me a while to get back to this as my understanding is
> currently very limited.  Poke me if I've not replied in a few weeks.
Sorry, still not had the spare time to take a proper look at this.
I'll try and grab some time during the week.

Jonathan

> 
> Thanks,
> 
> Jonathan
> 
>