x86-64: __pa_symbol() vs. __pa()

x86-64: __pa_symbol() vs. __pa()

[Jan Beulich]

Can any of you recall the reason for the comment accompanying the
__pa_symbol() definition:

/* __pa_symbol should be used for C visible symbols.
   This seems to be the official gcc blessed way to do such arithmetic. */

To me this rather sounds like a compiler problem, and if it indeed
exists then at least a few incorrect use have slipped in recently:
- arch/x86/mm/ioremap.c passes bm_pte to __pa()
- arch/x86/mm/pageattr.c passes _text and __{start,end}_rodata
  to __pa()
- arch/x86/mm/pgtable.c and arch/x86/xen/smp.c pass
  swapper_pg_dir to __pa()
- arch/x86/xen/enlighten passes level3_user_vsyscall to __pa()
This is only the set where x86-64 (which prior to the merge had
above comment already) is affected, i386 would have a few more.

What would be particularly problematic is the potential indirect use
of __pa(), e.g. through Xen's virt_to_{machine,bus}() as well as
the use in common code (e.g. kernel/kexec.c).

Thanks, Jan

Re: x86-64: __pa_symbol() vs. __pa()

[Andi Kleen]

On Tue, Dec 09, 2008 at 10:05:02AM +0000, Jan Beulich wrote:
> Can any of you recall the reason for the comment accompanying the
> __pa_symbol() definition:
> 
> /* __pa_symbol should be used for C visible symbols.
>    This seems to be the official gcc blessed way to do such arithmetic. */

gcc assumes that such symbol arithmetic doesn't wrap. iirc there was a 
miscompilation on PPC64 because of that and then it was fixed
everywhere. On x86-64 there are normally no wraps, but there might
be in some cases.

-Andi