Re: [PATCH RFC 1/6] ARM: pxa: Convert Spitz OHCI to GPIO descriptors

[Andy Shevchenko <andy@kernel.org>]

On Sun, Oct 01, 2023 at 11:18:41AM +0300, Andy Shevchenko wrote:
> On Wed, Sep 27, 2023 at 04:01:58PM +0200, Linus Walleij wrote:
> > On Mon, Sep 25, 2023 at 9:30 AM Andy Shevchenko <andy@kernel.org> wrote:

...

> > > > +     if (pxa_ohci->usb_host)
> > > > +             gpiod_put(pxa_ohci->usb_host);
> > >
> > > Linus, Bart, do we have misdesigned _optinal() GPIO APIs?
> > >
> > > In GPIOLIB=n, the above requires that redundant check. Shouldn't we replace
> > > gpiod_put() stub to be simply no-op?
> > 
> > You mean the WARN_ON(desc) in gpiod_put() in the static inline
> > stub version?
> > 
> > I thought about it for a bit, drafted a patch removing them, and then
> > realized the following:
> > 
> > If someone is making the gpiolib optional for a driver, i.e. neither
> > DEPENDS ON GPIOLIB nor SELECT GPIOLIB, they are a quite
> > narrow segment. I would say in 9 cases out of 10 or more this is
> > just a driver that should depend on or select GPIOLIB.
> > 
> > I think such drivers should actually do the NULL checks and not be
> > too convenient, the reason is readability: someone reading that
> > driver will be thinking gpios are not optional if they can call
> > gpiod_set_value(), gpiod_put() etc without any sign that the
> > desc is optional.
> > 
> > If the driver uses [devm_]gpiod_get_optional() the library is not
> > using the stubs and does the right thing, and it is clear that
> > the GPIO is *runtime* optional.
> > 
> > But *compile time* optional, *combined* with runtime optional -
> > I'm not so happy if we try to avoid warnings around that. I think
> > it leads to confusing configs and code that looks like gpiolib is
> > around despite it wasn't selected.
> > 
> > If the code isn't depending on or selecting GPIOLIB and still
> > use _optional() calls, it better be ready to do some extra checks,
> > because this is a weird combo, it can't be common.
> > 
> > Could be a documentation update making this clear though.
> > 
> > What do you other people think?
> 
> The problem here indeed if the code is not selecting or being dependent on
> GPIOLIB and uses _optional() calls.
> 
> I agree that this is quite a niche that should be addressed on the driver side.

One more thing, though. I think those warnings are incomplete or actually
reversed, and we outta use WARN_ON(IS_ERR(desc)), no?

This way it will fix my concerns and your concerns will be satisfied, right?
So, if gpiod_get() returns an error pointer and then we are trying to
free it with GPIOLIB=n, _then_ we will got a warning and it's obvious that
driver has to be prepared for that, otherwise if we have it NULL and
call for gpiod_get_optional(), even with GPIOLIB=n, it's fine to free, we
don't care.

-- 
With Best Regards,
Andy Shevchenko