Message-ID: <Z0YpAMz4RlWwc9Zq@pollux>
Date: Tue, 26 Nov 2024 21:01:04 +0100
From: Danilo Krummrich <dakr@...nel.org>
To: Rob Herring <robh@...nel.org>
Cc: gregkh@...uxfoundation.org, rafael@...nel.org, bhelgaas@...gle.com,
	ojeda@...nel.org, alex.gaynor@...il.com, boqun.feng@...il.com,
	gary@...yguo.net, bjorn3_gh@...tonmail.com, benno.lossin@...ton.me,
	tmgross@...ch.edu, a.hindborg@...sung.com, aliceryhl@...gle.com,
	airlied@...il.com, fujita.tomonori@...il.com, lina@...hilina.net,
	pstanner@...hat.com, ajanulgu@...hat.com, lyude@...hat.com,
	daniel.almeida@...labora.com, saravanak@...gle.com,
	rust-for-linux@...r.kernel.org, linux-kernel@...r.kernel.org,
	linux-pci@...r.kernel.org, devicetree@...r.kernel.org
Subject: Re: [PATCH v3 15/16] rust: platform: add basic platform device /
 driver abstractions

On Tue, Nov 26, 2024 at 01:15:59PM -0600, Rob Herring wrote:
> On Tue, Nov 26, 2024 at 9:17 AM Danilo Krummrich <dakr@...nel.org> wrote:
> >
> > On Tue, Nov 26, 2024 at 08:44:19AM -0600, Rob Herring wrote:
> > > > > > > The DT type and name fields are pretty much legacy, so I don't think the
> > > > > > > rust bindings need to worry about them until someone converts Sparc and
> > > > > > > PowerMac drivers to rust (i.e. never).
> > > > > > >
> > > > > > > I would guess the PCI cases might be questionable, too. Like DT, drivers
> > > > > > > may be accessing the table fields, but that's not best practice. All the
> > > > > > > match fields are stored in pci_dev, so why get them from the match
> > > > > > > table?
> > > > > >
> > > > > > Fair question, I'd like to forward it to Greg. IIRC, he explicitly requested to
> > > > > > make the corresponding struct pci_device_id available in probe() at Kangrejos.
> > >
> > > Making it available is not necessarily the same thing as passing it in
> > > via probe.
> >
> > IIRC, that was exactly the request.
> >
> > > I agree it may need to be available in probe(), but that
> > > can be an explicit call to get it.
> >
> > Sure, I did exactly that for the platform abstraction, because there we may
> > probe through different ID tables.
> 
> TBC, I think of_match_device() (both calling the C API and the method)
> should not be part of this series. I think we agreed on that already.
> Only if there is a need at some point later should we add it.

That matches my understanding.

> 
> > A `struct pci_driver`'s probe function has the following signature [1] though:
> >
> > `int (*probe)(struct pci_dev *dev, const struct pci_device_id *id)`
> >
> > [1] https://elixir.bootlin.com/linux/v6.12/source/include/linux/pci.h#L950
> 
> We have a mixture of probe with and without the _device_id parameter.
> I'd question if we really want to keep that for PCI when we have a
> chance to align things with Rust. We can't really with C as it would
> be too many drivers to change. Passing the _device_id only works if
> firmware matching is never used which can change over time. But if
> aligning things is not something we want to do, then I'll shut up.

I don't disagree. Again, this one is on Greg to comment on. Personally, I'm
fine with both.

> 
> > > > > Which table gets passed in though? Is the IdInfo parameter generic and
> > > > > can be platform_device_id, of_device_id or acpi_device_id? Not sure if
> > > > > that's possible in rust or not.
> > > >
> > > > Not sure I can follow you here.
> > > >
> > > > The `IdInfo` parameter is of a type given by the driver for driver specific data
> > > > for a certain ID table entry.
> > > >
> > > > It's analogue to resolving `pci_device_id::driver_data` in C.
> > >
> > > As I said below, the PCI case is simpler than for platform devices.
> > > Platform devices have 3 possible match tables. The *_device_id type we
> > > end up with is determined at runtime (because matching is done at
> > > runtime), so IdInfo could be any of those 3 types.
> >
> > `IdInfo` is *not* any of the three *_device_id types. It's the type of the
> > drivers private data associated with an entry of any of the three ID tables.
> 
> Ah yes, indeed. So no issue with the probe method.
> 
> > It is true that a driver, which registers multiple out of those three tables is
> > currently forced to have the same private data type for all of them.
> 
> I think that's a feature actually as it enforces best practices.

Agreed.

> 
> > I don't think this is a concern, is it? If so, it's easily resolvable by just
> > adding two more associated types, e.g. `PlatformIdInfo`, `DtIdInfo` and
> > `AcpiIdInfo`.
> >
> > In this case we would indeed need accessor functions like `dt_match_data`,
> > `platform_match_data`, `acpi_match_data`, since we don't know the type at
> > compile time anymore.
> 
> Do we need to split those out in rust or can we just call
> device_get_match_data()?

We'd need to split them, because they potentially would return different types.
(Again, I don't think that's necessary though.)

For C it's just a void pointer, so we don't bother there, but in Rust we'd want
it type safe.

> 
> >
> > I don't think that's necessary though.
> 
> Even if you don't support all 3 tables now, at a minimum I think you
> need to rename things to be clear what table type is supported and
> allow for adding the other types. For example, T::ID_TABLE needs to be
> renamed to be clear it's the of_device_id table.

Sure, I already got this on my list of changes for the next version.

> 
> Rob
> 

Powered by blists - more mailing lists