Re: [PATCH] per-cpu areas for 2.5.3-pre6

[Rusty Russell]

In message <[e-mail]> you write:
> Have we already forgotten the ppc reloc flamefest?  Better
> written as
>
> #define per_cpu(var, cpu)					\
>   ({ __typeof__(&(var)) __ptr;					\
>      __asm__ ("" : "=g"(__ptr)					\
> 	      : "0"((void *)&(var) + per_cpu_offset(cpu)));	\
>      *__ptr; })

"better".  Believe me, I was fully aware, but I refuse to write such
crap unless *proven* to be required.

> > +/* Created by linker magic */
> > +extern char __per_cpu_start, __per_cpu_end;
> [...]
> > +	per_cpu_size = ((&__per_cpu_end - &__per_cpu_start) + PAGE_SIZE-1)
> 
> Will fail on targets (e.g. alpha and mips) that have a notion of a
> "small data area" that can be addressed with special relocs.
> 
> Better written as
> 
>   extern char __per_cpu_start[], __per_cpu_end[];
>   per_cpu_size = (__per_cpu_end - __per_cpu_start) ...

I agree that this is much better.  But do not understand what small
relocs have to do with simple address arithmetic?  You've always been
right before: what am I missing?

Rusty.
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.