Re: [PATCH v3 0/8] iio: add new backend framework

["Nuno Sá" <noname.nuno@gmail.com>]

On Thu, 2023-12-14 at 08:16 -0600, Rob Herring wrote:
> On Wed, Dec 13, 2023 at 04:02:31PM +0100, Nuno Sa wrote:
> > v1:
> >  
> > https://lore.kernel.org/linux-iio/20231204144925.4fe9922f@jic23-huawei/T/#m222f517
> > 5273b81dbfe40b7f0daffcdc67d6cb8ff
> > 
> > v2:
> >  https://lore.kernel.org/r/20231208-dev-iio-backend-v2-0-5450951895e1@analog.com
> > 
> > Changes in v3:
> > - Patch 1:
> >  * Use proposed generic schema [1]. Also make it a required property;
> >  * Improved the commit message.
> > - Patch 2:
> >  * Improved commit message.
> > - Patch 4:
> >  * Namespace all IIO DMAENGINE buffer exports;
> >  * Removed unrelated new line removal change.
> > - Patch 5:
> >  * Namespace all IIO backend exports.
> > - Patch 6:
> >  * Set backend.h in alphabetical order;
> >  * Import IIO backend namespace.
> > - Patch 7:
> >  * Don't depend on OF in kbuild anymore;
> >  * Import IIO backend namespace.
> > 
> > For the bindings patches, I tried not to enter into much details about
> > the IIO framework as I think specifics of the implementation don't care
> > from the bindings perspective. Hopefully the commit messages are good
> > enough.
> > 
> > I'm also aware that patch 1 is not backward compatible but we are
> > anyways doing it on the driver side (and on the driver the property is
> > indeed required). Anyways, just let me know if making the property
> > required is not acceptable (I'm fairly confident no one was using the
> > upstream version of the driver and so validating devicetrees for it). 
> > 
> > Keeping the block diagram in v3's cover so we don't have to follow links
> > to check the one of the typicals setups. 
> > 
> >                                            --------------------------------------
> > -----------------
> >  ------------------                        | -----------         ------------
> >       -------  FPGA |
> >  |     ADC        |------------------------| | AXI ADC |---------| DMA CORE |----
> > --| RAM |       |
> >  | (Frontend/IIO) | Serial Data (eg: LVDS) | |(backend)|---------|          |----
> > --|     |       |
> >  |                |------------------------| -----------         ------------
> >       -------       |
> >  ------------------                        --------------------------------------
> > -----------------
> 
> Why doesn't axi-adc just have an io-channels property to adc? It's the 
> opposite direction for the link, but it seems more logical to me that 
> axi-adc depends on adc rather than the other way around.
> 

We are not interested on a single channel but on the complete device. From the axi-
adc point of view, there's not much we could do with the adc channel. I mean, maybe
we could still do something but it would be far from ideal (see below).

The opposite direction is exactly what we had (look at patch 2) just with another
custom property. The problem with that is that we would then need a bidirectional
link (we would need to callback into the provider and the provider would need to
access the consumer) between consumers and providers and that would be far from
optimal. The bidirectional link would exist because if we want to support fundamental
things like LVDS/CMOS interface tuning we need to set/get settings from the axi-adc
core. And as Jonathan suggested, the real data is captured or sent on the converters
(ADC or DACs) and that is why we have the IIO device/interface in there and why we
call them "frontends". In ADI usecases, backends are these FPGA cores providing
"services" to the "real" high speed converters. To put it in another way, the real
converter is the one who knows how to use the axi-adc core and not the other way
around. That's also one of the reasons why it would be way more difficult to handle
things with the opposite link. That's how we have done it so far and the mess we have
out of tree is massive :sweat_smile: We ended up doing raw writes and reads on the
axi-adc MMIO registers from the converter driver just because we had to configure or
get something from the axi-adc device but the link was backwards.

> And if there's another consumer in the chain, then a node could 
> certainly be both an io-channels consumer and producer.
> 

This should also be possible with this architecture. A node can be both a backend
(provider) and a consumer and we have an out of tree design that fits this (that I
surely want to upstream after the foundations are done).

> The architecture of the drivers seems odd to me. It looks similar to 
> making a phy driver handle all the state and protocol with the host 
> controller being a backend.

In this case, it's not really a controller. It's more like an extension of the device
because we need a way to handle the high sample rates this ADCs can do. Then we can
also do test tones with the backend which is useful for interface tuning (as
mentioned above).

To give you a bit more context, I'm adding the generic property because we will have
more users for it (from ADI - the next should be the axi-dac core) but STM is also
interested in this (likely the next user).

Hope this makes it a bit more clear...

- Nuno Sá