Re: [PATCH v2 2/2] drivers: misc: adi-axi-tdd: Add TDD engine

["Arnd Bergmann" <arnd@arndb.de>]

On Thu, Sep 28, 2023, at 06:54, Balas, Eliza wrote:
>> <conor+dt@kernel.org>; derek.kiernan@amd.com; dragan.cvetic@amd.com; Greg Kroah-Hartman <gregkh@linuxfoundation.org>;
>> linux-kernel@vger.kernel.org; devicetree@vger.kernel.org
>> Subject: Re: [PATCH v2 2/2] drivers: misc: adi-axi-tdd: Add TDD engine
>> 
>> [External]
>> 
>> On Thu, Sep 28, 2023, at 11:28, Eliza Balas wrote:
>> > This patch introduces the driver for the new ADI TDD engine HDL.
>> > The generic TDD controller is in essence a waveform generator
>> > capable of addressing RF applications which require Time Division
>> > Duplexing, as well as controlling other modules of general
>> > applications through its dedicated 32 channel outputs.
>> >
>> > The reason of creating the generic TDD controller was to reduce
>> > the naming confusion around the existing repurposed TDD core
>> > built for AD9361, as well as expanding its number of output
>> > channels for systems which require more than six controlling signals.
>> >
>> > Signed-off-by: Eliza Balas <eliza.balas@analog.com>
>> 
>> Thanks for your submission, I've had a first look at the driver
>> and the implementation of the interface you have chosen looks
>> all good to me, so I have no detailed comments on that.
>> 
>> It would however help to explain the ideas you had for the
>> user-space interface design and summarize them in the changelog
>> text.
>> 
>> You have chosen a low-level interface that wraps the individual
>> device registers and gives user space direct control over them.
>> The risk here is to lock yourself into the first design,
>> giving you less flexibility for future extensions, so it would
>> help to understand what the usage model is here.
>> 
>> One risk is that there may be an in-kernel user in the future
>> when the TDD engine interacts with another device, so you
>> need a driver level interface, which would in turn break
>> if any user pokes the registers directly.
>> 
>> Another possible problem I see is that an application written
>> for this driver would be incompatible with similar hardware
>> that has the same functionality but a different register-level
>> interface, or even a minor revision of the device that ends up
>> breaking one of the assumptions about the hardware design.
>> 
>> In both cases, the likely answer is to have a higher-level
>> interface of some sort, but the downside of that would be
>> that it is much harder to come up with a good interface that
>> covers all possible use cases.
>> 
>> Another question is whether you could fit into some
>> existing subsystem instead of creating a single-driver
>> interface. drivers/iio/ might be a good choice, as
>> it already handles both in-kernel and userspace users,
>> and provides a common abstraction for multiple classes
>> of devices that (without any domain knowledge in my case)
>> look similar enough that this could be added there.
>> 
>
> We are using this driver with an iio-fake device 
> https://github.com/analogdevicesinc/linux/blob/master/Documentation/devicetree/bindings/iio/jesd204/adi%2Ciio-fakedev.yaml 
>  so we can take advantage of the iio user-space interface.

I don't understand how that works yet: Do you mean that there
is  user-space application that uses the tdd sysfs interface to
export an IIO device back into the kernel, or do you mean there
is a regular IIO device in with a kernel driver that is used
as the back-end for the tdd device, or something else?

> We talked in the previous v1 patch emails about adding this driver to 
> an existing subsystem, and I raised the question if we should add it to 
> the iio subsystem, but the driver is not registered into the IIO device 
> tree, and does not rely on IIO kernel APIs, so I concluded that misc is 
> a better choice.
> What do you think?

My feeling is that if you can make it fit into IIO, then this is
likely the better choice, unless you can guarantee that this is
a one-off driver with a single hardware implementation and
a single userspace. If you need the flexibility later to do
more things, the risk is that you end up duplicating a lot of
functionality that already exists in IIO.

This would of course mean using the interfaces provided by the
IIO core, with the addition of a tdd device type rather than
just having a standalone driver with just the sysfs interface
you have here.

      Arnd