Re: [PATCH] reset: allow to pass NULL pointer to reset_control_put()

[Arnd Bergmann]

On Wednesday 04 May 2016 14:34:43 Philipp Zabel wrote:
> Am Mittwoch, den 04.05.2016, 20:34 +0900 schrieb Masahiro Yamada:
> > Hi Arnd,
> > 
> > 2016-05-04 20:24 GMT+09:00 Arnd Bergmann <arnd@xxxxxxxx>:
> > > On Wednesday 04 May 2016 20:17:51 Masahiro Yamada wrote:
> > >> Currently, reset_control_put() just returns for error pointer,
> > >> but not for NULL pointer.  This is not reasonable.
> > >>
> > >> Passing NULL pointer should be allowed as well to make failure path
> > >> handling easier.
> > >>
> > >> Signed-off-by: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> > >> ---
> > >>
> > >>  drivers/reset/core.c | 2 +-
> > >>  1 file changed, 1 insertion(+), 1 deletion(-)
> > >>
> > >> diff --git a/drivers/reset/core.c b/drivers/reset/core.c
> > >> index 181b05d..7bb16d1 100644
> > >> --- a/drivers/reset/core.c
> > >> +++ b/drivers/reset/core.c
> > >> @@ -288,7 +288,7 @@ EXPORT_SYMBOL_GPL(reset_control_get);
> > >>
> > >>  void reset_control_put(struct reset_control *rstc)
> > >>  {
> > >> -       if (IS_ERR(rstc))
> > >> +       if (IS_ERR_OR_NULL(rstc))
> > >>                 return;
> > >>
> > >>         module_put(rstc->rcdev->owner);
> > >
> > > Using IS_ERR_OR_NULL() normally indicates that there is something
> > > wrong with the API, or with the caller.
> > >
> > > What exactly is the idea behind treating an error pointer as a valid
> > > input to reset_control_put() here? Maybe it should just test for
> > > NULL?
> 
> The idea was that you could do
>         drvdata->rstc = reset_control_get(...)
> in the probe() function and
>         reset_control_put(drvdata->rstc)
> in remove() without having to check for IS_ERR(drvdata->rstc) again.
> I'm not convinced this is necessarily a good idea though. To simplify
> the teardown path we already have  devm_reset_control_get().
> 
> > I thought about that a bit,
> > but there might be some (not nice) drivers that rely on the current behavior.
> > I did not want to break any boards with my patch.
> > 
> > So, should it be
> > 
> >           if (!rstc)
> >                     return;
> > or, perhaps
> > 
> >           if (!rstc || WARN_ON_ONCE(IS_ERR(rstc)))
> >                     return;
> > 
> > ?
> 
> NULL is not a valid input to reset_control, reset_control_get(_optional)
> should never return NULL.
> 
> I'd be in favor of turning this into
> 
>         if (WARN_ON(IS_ERR_OR_NULL(rstc)))
>                 return;

Sounds good to me too. We'd still have to think about whatever
Masahiro was trying to do and how his caller should be written,
but hopefully there is a good solution.

> As far as I am aware, ehci-tegra is the only driver that currently makes
> use of the IS_ERR(rstc) return in reset_control_put():
> 
>         struct reset_control *usb1_reset;
> 
>         usb1_reset = of_reset_control_get(phy_np, "usb");
>         if (IS_ERR(usb1_reset)) {
>                 /* ... */
>         } else {
>                 reset_control_assert(usb1_reset);
>                 udelay(1);
>                 reset_control_deassert(usb1_reset);
>         }
>         reset_control_put(usb1_reset);
> 
> That'd be trivial to fix.

Ah, good. Could the above code just be converted into a variation
of device_reset() that takes the name of a reset line?

	Arnd