Re: [PATCH] mm: disallow early_pfn_to_nid on configurations which do not implement it

[Michal Hocko]

On Wed 05-07-17 16:00:55, Andrew Morton wrote:
> On Tue,  4 Jul 2017 09:58:03 +0200 Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> 
> > From: Michal Hocko <mhocko@xxxxxxxx>
> > 
> > early_pfn_to_nid will return node 0 if both HAVE_ARCH_EARLY_PFN_TO_NID
> > and HAVE_MEMBLOCK_NODE_MAP are disabled. It seems we are safe now
> > because all architectures which support NUMA define one of them (with an
> > exception of alpha which however has CONFIG_NUMA marked as broken) so
> > this works as expected. It can get silently and subtly broken too
> > easily, though. Make sure we fail the compilation if NUMA is enabled and
> > there is no proper implementation for this function. If that ever
> > happens we know that either the specific configuration is invalid
> > and the fix should either disable NUMA or enable one of the above
> > configs.
> > 
> > ...
> >
> > --- a/include/linux/mmzone.h
> > +++ b/include/linux/mmzone.h
> > @@ -1055,6 +1055,7 @@ static inline struct zoneref *first_zones_zonelist(struct zonelist *zonelist,
> >  	!defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
> >  static inline unsigned long early_pfn_to_nid(unsigned long pfn)
> >  {
> > +	BUILD_BUG_ON(IS_ENABLED(CONFIG_NUMA));
> >  	return 0;
> >  }
> >  #endif
> 
> Wouldn't this be more conventional?

Well, both would lead to a compilation errors which is what I want to
achieve. The above is easier to parse IMHO. If you believe a longer
ifdef chain is better I won't object.

> --- a/include/linux/mmzone.h~a
> +++ a/include/linux/mmzone.h
> @@ -1052,7 +1052,8 @@ static inline struct zoneref *first_zone
>  #endif
>  
>  #if !defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) && \
> -	!defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
> +	!defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) && \
> +	!defined(CONFIG_NUMA)
>  static inline unsigned long early_pfn_to_nid(unsigned long pfn)
>  {
>  	return 0;
> _
> 

-- 
Michal Hocko
SUSE Labs