Re: [PATCH -next] misc: rtsx: fix build for CONFIG_PM not set

[Jonathan Cameron <Jonathan.Cameron@Huawei.com>]

On Sun, 6 Mar 2022 15:29:17 +0000
Paul Cercueil <paul@crapouillou.net> wrote:

> Hi Jonathan,
> 
> Le dim., mars 6 2022 at 15:12:12 +0000, Jonathan Cameron 
> <Jonathan.Cameron@Huawei.com> a écrit :
> > On Sun, 27 Feb 2022 17:56:31 +0000
> > Paul Cercueil <paul@crapouillou.net> wrote:
> >   
> >>  Le dim., févr. 27 2022 at 18:51:38 +0100, Arnd Bergmann
> >>  <arnd@arndb.de> a écrit :  
> >>  > On Sun, Feb 27, 2022 at 6:46 PM Paul Cercueil   
> >> <paul@crapouillou.net>  
> >>  > wrote:  
> >>  >>  Le dim., févr. 27 2022 at 18:30:16 +0100, Arnd Bergmann
> >>  >>
> >>  >>  There could be a DEFINE_DEV_PM_OPS(), but I don't think that's
> >>  >> really
> >>  >>  needed - you can very well declare your struct dev_pm_ops   
> >> without  
> >>  >> using
> >>  >>  one of these macros. Just make sure to use the   
> >> SYSTEM_SLEEP_PM_OPS /  
> >>  >>  RUNTIME_PM_OPS macros for the callbacks and pm_ptr() for the
> >>  >> device.pm
> >>  >>  pointer.  
> >>  >
> >>  > Ah, of course, so it comes down to
> >>  > s/SET_SYSTEM_SLEEP_PM_OPS/SYSTEM_SLEEP_PM_OPS/ while
> >>  > removing all the #ifdef an __maybe_unused annotations. The   
> >> pm_ptr()  
> >>  > in driver.pm makes this slightly more optimized AFAICT, but has no
> >>  > effect on behavior, right?  
> >> 
> >>  The use of SYSTEM_SLEEP_PM_OPS makes sure that the callbacks are
> >>  dropped if the dev_pm_ops is dead code, and the pm_ptr() must be 
> >> used
> >>  for the compiler to know that the dev_pm_ops is dead code.
> >> 
> >>  -Paul
> >> 
> >>   
> > 
> > Hi Paul,
> > 
> > We have one remaining case which is still ugly to do.
> > Where both SYSTEM_SLEEP_PM_OPS/RUNTIME_PM_OPS are set and
> > the dev_pm_ops structure is exported.
> > 
> > For that one we still need to expose #ifdef fun in the
> > drivers I think.
> > 
> > Any suggestions on a clean solution for that?  
> 
> Use the _EXPORT_DEV_PM_OPS() macro?
> 
> If you make it call __EXPORT_SYMBOL() (with two underscores instead of 
> one) you can specify the namespace as well. All you need then is a nice 
> wrapper macro in pm_runtime.h, that can be used in the driver.

Yup. The patch for adding namespace versions in the first place
did that.

https://lore.kernel.org/linux-iio/20220220181522.541718-1-jic23@kernel.org/T/#m5a3dc798606f88c10870a66e18b5b175e1a30243

I was just hoping we could maybe avoid adding a few more macros, particularly
ones that take lots of arguments like this one will.

We'll have something like 4 more macros

#define EXPORT_NS_GPL_DEV_PM_OPS(name, suspend_fn, resume_fn, runtime_suspend_fn, \
				 runtime_resume_fn, idle_fn, sec, ns) \
	_EXPORT_DEV_PM_OPS(name, suspend_fn, resume_fn, runtime_suspend_fn, \
			   runtime_resume_fn, idle_fn, "_gpl", ns)
#define EXPORT_NS_DEV_PM_OPS() ...
#define EXPORT_DEV_PM_OPS() ...
#define EXPORT_GPL_DEV_PM_OPS() ...

if we provide the same for non NS cases.

I'll roll a patch for next cycle hopefully building on the one referenced
above if Rafael picks that up for this merge window.

Jonathan



> 
> Cheers,
> -Paul
> 
> > Currently I have this...
> > 
> > #ifdef CONFIG_PM
> > const struct dev_pm_ops bmc150_magn_pm_ops = {
> > 	SYSTEM_SLEEP_PM_OPS(...)
> > 	RUNTIME_PM_OPS(...)
> > };
> > EXPORT_SYMBOL_NS(bmc150_magn_pm_ops, IIO_BMC150_MAGN);
> > #else
> > static const __maybe_unused dev_pm_ops bmc150_magn_pm_ops = {
> > 	SYSTEM_SLEEP_PM_OPS(...)
> > 	RUNTIME_PM_OPS(...)
> > };
> > #endif
> > Not super clean but perhaps we do need
> > EXPORT_NS_DEV_PM_OPS
> > EXPORT_NS_GPL_DEV_PM_OPS
> > and potentially the non namespaced versions.
> > 
> > Thanks,
> > 
> > Jonathan  
> 
>