Why do we need reset_control_get_optional() ?

[Philipp Zabel]

Am Mittwoch, den 24.08.2016, 15:58 +0900 schrieb Masahiro Yamada:
> Hi Philipp,
> 
> 2016-08-16 23:36 GMT+09:00 Masahiro Yamada <yamada.masahiro at socionext.com>:
> > Hi Philipp, Arnd.
> >
> >
> >
> > 2016-08-09 1:39 GMT+09:00 Philipp Zabel <p.zabel at pengutronix.de>:
> >> Am Freitag, den 05.08.2016, 17:50 +0200 schrieb Arnd Bergmann:
> >>> On Thursday, July 28, 2016 1:00:49 PM CEST Philipp Zabel wrote:
> >>> > Am Donnerstag, den 28.07.2016, 19:52 +0900 schrieb Masahiro Yamada:
> >>>
> >>> > > > In my experimental patch, I make the _optional functions
> >>> > > > return NULL if no "resets" property is provided but return
> >>> > > > an error if there are reset lines but the subsystem is
> >>> > > > disabled, i.e. an optional reset must be used if it's in the
> >>> > > > DT, but can be ignored otherwise.
> >>> > >
> >>> > > I do not like this idea.
> >>> > >
> >>> > > reset_control_get() (or variants) should not return NULL, it is ambiguous.
> >>> > > It should return ERR_PTR(-ENOENT) if no "resets" property.
> >>> > >
> >>> > > I only want two types for functions that return a pointer.
> >>> > >
> >>> > > [1] return a valid pointer on success, or return NULL on failure
> >>> > >     (for example, kmalloc())
> >>> > > [2] return a valid pointer on success, or return error pointer on failure
> >>> > >    (many of _register() functions)
> >>> > >
> >>> > > Mixing [1] and [2] will be a mess.
> >>>
> >>> Ah, right. I was thinking only of the case where the reset subsystem
> >>> is completely disabled here, so returning NULL could be considered
> >>> a valid return code that can in turn be passed into the other
> >>> functions.
> >>>
> >>> However, I agree that returning NULL as a valid result from
> >>> ..._get_optional() would be bad style, so let's drop my idea
> >>> there.
> >>>
> >>> > I too would prefer to keep that as-is. The reset_control_get_optional
> >>> > stub could return -ENOENT if there is no resets device tree property.
> >>>
> >>> Now I'm also confused about what we really need
> >>> reset_control_get_optional() for, and which error codes the callers
> >>> are supposed to check.
> >>>
> >>> This is the matrix I think you mean for _get_optional:
> >>>
> >> [...]
> >>> CONFIG_RESET_CONTROLLER=n, dt entry present: -EOPNOTSUPP
> >>> CONFIG_RESET_CONTROLLER=n, dt entry missing: -ENOENT
> >>
> >> ^^ I didn't consider this distiction.
> >>
> >>> Is this what you had in mind? If so, what is the value of the
> >>> added runtime warning for reset_control_get? Any caller of that
> >>> function would already check for errors, the only difference
> >>> I see is that callers of _optional can ignore -ENOENT.
> >>
> >> My initial motivation was to make it as hard as possible to misconfigure
> >> the kernel, which is why I initially didn't want stubs for the
> >> non-optional variant. Of course that would cause build failures and/or
> >> reduced compile test coverage, so we added the stubs and the warning to
> >> make it obvious when a misconfigured kernel is running: on a kernel with
> >> RESET_CONTROLLER=n drivers that use reset_control_get are expected to
> >> build, but they are not expected to work. I suppose the same is the case
> >> for _optional, if the dt entry is present, so maybe we should drop
> >> reset_control_get_optional and add always a warning in case of
> >> -EOPNOTSUPP.
> >> I don't want all drivers to have to differentiate between -EOPNOTSUPP
> >> and -ENOENT error codes, only current reset_control_get_optional users
> >> have to do that.
> >
> >
> > I've posted a patch to drop reset_control_get_optional;
> > https://patchwork.kernel.org/patch/9284063/
> >
> > Could you check if it works?
> >
> > If we go this way, my patch
> > 289363fd99a17d6249ee1373541f1da43cbb22c5
> > in your reset/next branch is completely useless.
> >
> > As the commits in the reset-subsystem do not appear
> > even in linux-next until they are pulled into the ASOC tree,
> > how about dropping 289363fd and turning around?
> >
> 
> 
> If you want to take time for this topic,
> how about dropping 289363fd99a17d6249ee1373541f1da43cbb22c5
> ("reset: add WARN_ON(1) to non-optional reset_control_get variants")
> for now?
> 
> 
> I noticed some reset consumers already started dropping _optional,
> while their reset lines should be really optional.
> 
> See
> commit d0e08b0077f49e209bc90305ddf1ca434ac6cc0e
> commit 62d9694a003dba585026df36c181e3ca930aeafc
> 
> Even generic drivers such as ehci-platform.c / ohci-platform.c
> opted out of _optional.
> 
> 
> If commit 289363fd99a17d6249ee1373541f1da43cbb22c5 is merged,
> users of the generic drivers but without reset-controller
> will start to complain about the WARN_ON(1)  sooner or later.

Hmm, I would really like to keep the warnings, but as Arnd suggested,
only if the device tree property is present. I'll drop your patch for
now.

regards
Philipp