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

[Jonathan Cameron]

On Sun, 27 Feb 2022 17:56:31 +0000
Paul Cercueil <paul@xxxxxxxxxxxxxxx> wrote:

> Le dim., févr. 27 2022 at 18:51:38 +0100, Arnd Bergmann 
> <arnd@xxxxxxxx> a écrit :
> > On Sun, Feb 27, 2022 at 6:46 PM Paul Cercueil <paul@xxxxxxxxxxxxxxx> 
> > 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?

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