On Wed, Aug 02, Andrew Morton wrote:

> > +#define per_cpu(var, cpu) (*({				\
> > +	int user_error_##var __attribute__ ((unused));	\
> > +	RELOC_HIDE(&per_cpu__##var, __per_cpu_offset[cpu]); }))
> 
> What's it do?  Forces a syntax error if `var' isn't a simple identifier?
> 
> Seems sane, although I'd check that the compiler doesn't accidentally
> waste a stack slot for that local.  Perhaps it's be safer to make
> it a non-existing function:
> 
> 	extern int user_error#var(void);
> 

Yes, that is even better. See attached patch with the modifications for my
major archs (s390, x86-64, generic). I did a run of 'make allmodconfig' with
the attached patch and it produced some errors. I'll send the patches in a
seperate mail.