Re: [PATCH net-next 07/10] net: phy: switch to using class_find_device_by_fwnode()

On Mon, Mar 23, 2026 at 06:39:28PM +0000, Russell King (Oracle) wrote:
> On Mon, Mar 23, 2026 at 11:33:36AM -0700, Dmitry Torokhov wrote:
> > On Mon, Mar 23, 2026 at 08:59:52AM +0000, Russell King (Oracle) wrote:
> > > On Sun, Mar 22, 2026 at 10:17:15PM -0700, Dmitry Torokhov wrote:
> > > > On Mon, Mar 23, 2026 at 03:54:09AM +0100, Andrew Lunn wrote:
> > > > > > -	d = class_find_device_by_of_node(&mdio_bus_class, mdio_bus_np);
> > > > > > +	d = class_find_device_by_fwnode(&mdio_bus_class,
> > > > > > +					of_fwnode_handle(mdio_bus_np));
> > > > > 
> > > > > When you look at this, why is it better?
> > > > 
> > > > I think we should move as much as possible towards firmware-agnostic
> > > > APIs and use fwnode_handle instead of device_node or software_node or
> > > > ACPI companion. To discourage this I think we better remove
> > > > firmware-specific APIs where we have firmware-agnostic ones and
> > > > eventually clean up drivers that use OF- or ACPI-specific APIs.
> > > 
> > > Basically, no. This is wrong.
> > > 
> > > It may sound like a good goal, but there's an underlying issue. This
> > > goal assumes that the firmware description in OF and ACPI are
> > > indentical.
> > 
> > If they are different then drivers will make allowance for this, like
> > I2C core or SPI core does. But most of the modern drivers use
> > firmware-agnostic APIs (device_property_*()).
> 
> What is appropriate is up to the standards bodies responsible for
> the firmware. ACPI has a separate body, and we can't just dump
> the structure we use for networking into ACPI. So no, you can't
> just switch to firmware-agnostic APIs for networking.

So I believe there is a difference between deciding:

1. What schema is being used on a particular system - it may be
OF-compatible, or ACPI compatible one, and ACPI may support
OF-compatible schemes for certain subsystems or individual devices, and

2. What is the API to access the properties.

There is no harm to use device_property_read_*() universally as well as 
not having OF-specific APIs at the driver core level. What matters is
schema matches the system type.

> 
> For infrastructure where we lookup stuff by some kind of firmware
> node, I have no problem with converting that to fwnode APIs,
> because that doesn't transfer the DT description into other
> firmwares without prior agreement of the appropriate firmware
> standards bodies.
> 
> This is not "foreign territory" - ACPI in general doesn't want to
> describe e.g. the individual components of a network card, unlike
> DT.

I think we actually broadly agree? What you are arguing against is
saying that OF schema is guaranteed to work on ACPI systems and vice
versa, but I am not saying that (although for some things it may). But
still most OF schema can be handled by generic APIs (either
device_property_read_*() or fwnode ones).

Thanks.

-- 
Dmitry