Hi Uwe,

On Thu, Mar 23, 2017 at 11:10 AM, Uwe Kleine-K?nig
[off-list ref] wrote:

On Thu, Mar 23, 2017 at 10:32:01AM +0100, Linus Walleij wrote:

quoted

On Mon, Mar 20, 2017 at 12:07 PM, Uwe Kleine-K?nig
[off-list ref] wrote:

quoted

On Mon, Mar 20, 2017 at 11:38:52AM +0100, Geert Uytterhoeven wrote:

quoted

On Mon, Mar 20, 2017 at 11:31 AM, Uwe Kleine-K?nig
[off-list ref] wrote:

quoted

On Mon, Mar 20, 2017 at 10:56:14AM +0100, Geert Uytterhoeven wrote:

quoted

Documentation/gpio/consumer.txt rightfully sates:
| Note that gpio_get*_optional() functions (and their managed variants), unlike
| the rest of gpiolib API, also return NULL when gpiolib support is disabled.
| This is helpful to driver authors, since they do not need to special case
| -ENOSYS return codes.  System integrators should however be careful to enable
| gpiolib on systems that need it.

I cannot find this paragraph in Documentation/gpio/consumer.txt:

        $ git grep -e 'ENOSYS' v4.11-rc3 -- Documentation/gpio/
        <void>

It was added by commit 22c403676dbbb7c6 ("gpio: return NULL from
gpiod_get_optional when GPIOLIB is disabled")

Ah, that's in next.

I still think this is wrong and I'm a bit disappointed that Linus merged
this patch (without saying so in the thread even) as I thought Linus and
I agreed on this being a bad idea.

I think it is not good, but what we have before this patch is worse.

I.e. it is the lesser evil.

Before this patch the API is inconsistent: it gives NULL if GPIOLIB
is defined and -ENOSYS if GPIOLIB is not defined.

So old is: it worked as intended with GPIOLIB and produced an error
without GPIOLIB even if no GPIO is there and NULL would be right.

quoted

After this patch it gives NULL in both cases, and that is at least
consistent.

And new is: with GPIOLIB it still works as intended and gives you NULL
even if a GPIO is there and so -ENOSYS would be right.

If I forget to enable a driver, the corresponding class device is also not
there.

I admit that most of the time with GPIOLIB NULL is the right answer
because probably there is no GPIO to handle. But if there is a GPIO you
run in hard to debug problems instead of being able to see the error
code and act accordingly.

But having the error breaks setups where the GPIO is optional and does
not exist.

Make sure to enable all drivers and subsystems you need when building
your kernel. That's always true. And may indeed be hard to debug (e.g. what
kernel options do I need to make systemd work?).

write(2) and close(2) succeed most of the time, too. Still it's not a
good idea to not check the return value. Or let the kernel return
success unconditionally.

Writing all bytes passed in the buffer is "optional" in another sense than
an "optional" GPIO: you must retry the write, while you can continue if
an optional GPIO is not present.

So you exchanged many obvious and easy to fix problems with a few hard
ones. I don't agree that's a good idea, but you seem to be willing to
try it. Good luck.

Yeah, before drivers had to explicitly ignore -ENOSYS if they want to
support platforms with and without GPIOLIB. Bad...

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert at linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds