Re: [PATCH v10 5/7] samples: rust: Add debugfs sample driver

[Matthew Maurer]

On Tue, Aug 19, 2025 at 5:34 PM Danilo Krummrich <dakr@xxxxxxxxxx> wrote:
>
> On Wed Aug 20, 2025 at 12:53 AM CEST, Matthew Maurer wrote:
> > Adds a new sample driver that demonstrates the debugfs APIs.
> >
> > The driver creates a directory in debugfs and populates it with a few
> > files:
> > - A read-only file that displays a fwnode property.
> > - A read-write file that exposes an atomic counter.
> > - A read-write file that exposes a custom struct.
> >
> > This sample serves as a basic example of how to use the `debugfs::Dir`
> > and `debugfs::File` APIs to create and manage debugfs entries.
> >
> > Signed-off-by: Matthew Maurer <mmaurer@xxxxxxxxxx>
>
> This is a great example, thanks! I really like how the API turned out.
>
> When it comes to the newly added Scope API - and I assume this does not come at
> a surprise - I have some concerns.

Yes, I expected this to be the case, but inspired by some of the
comments about wanting to just create files off fields and forget
about them, I wanted to take one more crack at it.

>
> But first, thanks a lot for posting the socinfo driver in both variants, with
> and without the Scope API.
>
> I had a brief look at both of those and I can see why you want this.
>
> With the Scope thing you can indeed write things a bit more compressed (I think
> in the patches the differences looks quite a bit bigger than it actually is,
> because the scope-based one uses quite some code from the file-based one).
>
> I think the downsides are mainly:
>
>   - The degree of complexity added for a rather specific use-case, that is also
>     perfectly representable with the file-based API.
I don't *think* this is just for this use case - if I just wanted to
improve the DebugFS use case, I'd mostly be looking at additional code
for `pin-init` (adding an `Option` placement + a few ergonomic
improvements to `pin_init` would slim off a large chunk of the code).
The idea here was that a file might not always directly correspond to
a field in a data structure, and the `File` API forces it to be one.
We could decide that forcing every file to be a data structure field
is a good idea, but I'm not certain it is.
>
>   - It makes it convinient to expose multiple fields grouped under the same lock
>     as separate files, which design wise we shouln't encourage for the reasons
>     we discussed in v8.
It's still pretty convenient to do this with `File`. I don't know how
common it'll be in kernel code, but in userspace Rust, `Arc<Mutex<T>>`
is a very common primitive. I would be unsurprised to see someone use
this pattern to expose separate fields as separate files if we go with
the `File` API.
>
> I think for the sake of getting this series merged, which I would really love to
> see, I think we should focus on the file-based API first. Once we got this
> landed I think we can still revisit the Scope idea and have some more discussion
> about it.

This is why I put the scope API and sample as patches on the end chain
of the series - it is possible to merge only the `File`-based API if
that's what we want to do first, and consider the rest later.

>
> I will have a more detailed look tomorrow (at least for the patches 1-5).
>
> Thanks again for working on this!
>
> - Danilo