Re: [RFC PATCH v2 07/16] gpio: Add support for unified device properties interface

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Friday, September 26, 2014 10:36:06 AM Arnd Bergmann wrote:
> On Thursday 25 September 2014 20:21:32 Darren Hart wrote:
> > On Wed, Sep 24, 2014 at 11:12:36AM +0200, Arnd Bergmann wrote:
> > > How would you feel about a more general way of probing LED, using
> > > a new helper in the leds-core that iterates over the child nodes
> > > and parses the standard properties but calls into a driver specific
> > > callback to parse the specific properties?
> > > It's probably much more work than your current approach, but it seems
> > > to me that there is more to gain by solving the problem for LED
> > > drivers in particular to cut down the per-driver duplication
> > > at the same time as the per-firmware-interface duplication.
> > > 
> > > As a start, we could probably take the proposed device_for_each_child_node
> > > and move that into the leds-core, changing the fw_dev_node argument
> > > for an led_classdev with the addition of the of_node and acpi_object
> > > members. It would still leave it up to the gpio-leds driver to do
> > > 
> > > 	if (led_cdev->of_node)
> > > 		gpiod = devm_of_get_gpiod(led_cdev->of_node, ...);
> > > 	else
> > > 		gpiod = devm_acpi_get_gpiod(led_cdev->acpi_object, ...);
> > 
> > So as Mika has pointed out, LEDs aren't the only ones affected. Several drivers
> > will need to walk through non-device child nodes, and it seems to me that having
> > a firmware-independent mechanism to do so benefits the drivers by both making
> > them smaller and by increasing the reusability of new drivers and drivers
> > updated to use the new API across platforms.
> > 
> > I fear we might be entering bike shed territory as we seem to be repeating
> > points now. Can you restate your concern with the interface and why this level
> > of abstraction is worse for the kernel? I'm not seeing this point, so I'm not
> > sure what to address in my response.
> 
> I think we should have abstractions for all common interfaces but make
> them as simple as possible. In the discussions at the kernel summit,
> everyone agreed that we should have common accessors for simple properties
> (bool, int, string, ...) based on device pointers, as well as subsystem
> specific accessors to handle the high-level abstractions (registers,
> interrupts, gpio, regulator, pinctrl, dma, reset, pwm, ...).
> 
> Having generalized accessors for the same properties in child nodes of
> the device goes beyond that, and I think this is the wrong trade-off
> between interface simplicity and generality since only few drivers will
> be able to use those. I think we will always have to live with a leaky
> abstraction because some drivers need to do things beyond what we can
> do with a common API.

Some drivers do, but then we can avoid adding DT/ACPI knowledge to some
drivers by adding general accessors for properties in child nodes.  In my
opinion, drivers should not do things specific to DT/ACPI unless that is
unavoidable.

> 
> > Grant, Linus W? Thoughts?
> 
> I definitely want to hear other voices on this too. This is really not
> a fundamental debate I think, but more a question of how far the abstraction
> should go.

Right.

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.