Re: [PATCH] x86: memtest: fix compile warning

[Andreas Herrmann <andreas.herrmann3@amd.com>]

On Thu, Jun 11, 2009 at 04:21:41PM +0200, Thomas Gleixner wrote:
> On Thu, 11 Jun 2009, Andreas Herrmann wrote:
> 
> > Commit c9690998ef48ffefeccb91c70a7739eebdea57f9
> > (x86: memtest: remove 64-bit division) introduced following compile warning:
> > 
> >  arch/x86/mm/memtest.c: In function 'memtest':
> >  arch/x86/mm/memtest.c:56: warning: comparison of distinct pointer types lacks a cast
> >  arch/x86/mm/memtest.c:58: warning: comparison of distinct pointer types lacks a cast
> > 
> > Signed-off-by: Andreas Herrmann <andreas.herrmann3@amd.com>
> > ---
> >  arch/x86/mm/memtest.c |    4 ++--
> >  1 files changed, 2 insertions(+), 2 deletions(-)
> > 
> > Sorry.
> > Please apply.
> 
> I applied it already, but zapped it right away, as it is bad style to
> do the type casting in the loops. The proper fix is below.

Doesn't your fix re-introduces the 64-bit division problem with old
gcc?  I removed that division with the mentioned commit just forgot to
type-cast the pointer.

> But aside of that this code is confusing.
> 
> 	start_phys_aligned = ALIGN(start_phys, incr);
>
>
> Why do we have to fiddle with the alignment. Are you really seing e820
> entries which are not 8 byte aligned ?

CC-ing Yinghai who might know more about this.
See also http://marc.info/?l=linux-kernel&m=123490434528131

> 	for (p = start; p < end; p++, start_phys_aligned += incr) {
> 		if (*p == pattern)
> 			continue;
> 		if (start_phys_aligned == last_bad + incr) {
> 			last_bad += incr;
> 			continue;
> 		}
> 		if (start_bad)
> 			reserve_bad_mem(pattern, start_bad, last_bad + incr);
> 		start_bad = last_bad = start_phys_aligned;
> 	}
> 	if (start_bad)
> 		reserve_bad_mem(pattern, start_bad, last_bad + incr);
> 
> I really had to look more than once to understand what the heck
> start_phys_aligned and last_bad + incr are doing. Really non
> intuitive.
>
> But the reserve_bad_mem() semantics are even more scary:
> 
> - if you hit flaky memory, which gives you bad and good results here
>   and there, you call reserve_bad_mem() totally unbound which is
>   likely to overflow the early reservation space and panics the
>   machine. You need to keep track of those events somehow (e.g. in a
>   bitmap) so you can detect such problems and mark the whole affected
>   region bad in one go.

Agreed, needs to be fixed.

> - you call reserve_early() which calls __reserve_early(....,
>   overrun_ok = 0) so if you do the default multi pattern scan and each
>   run sees the same region of broken memory you will trigger the
>   "Overlapping early reservations" panic in __reserve_early() when you
>   reserve that region the second time. Why do you run the test twice
>   when the first one failed already ? Also there is no need to do the
>   wipeout run in that case, which will trigger it as well!

Sure, needs to be fixed as well.

(Note: I think both problems exist in the memtest code right from the beginning.)

> So in both cases you panic the machine w/o need.
> 
> Please fix ASAP.

> Thanks,
> 
> 	tglx
> ---
> diff --git a/arch/x86/mm/memtest.c b/arch/x86/mm/memtest.c
> index d1c5cef..18d244f 100644
> --- a/arch/x86/mm/memtest.c
> +++ b/arch/x86/mm/memtest.c
> @@ -40,16 +40,14 @@ static void __init reserve_bad_mem(u64 pattern, u64 start_bad, u64 end_bad)
>  
>  static void __init memtest(u64 pattern, u64 start_phys, u64 size)
>  {
> -	u64 *p, *end;
> -	void *start;
> +	u64 *p, *start, *end;
>  	u64 start_bad, last_bad;
>  	u64 start_phys_aligned;
> -	size_t incr;
> +	const size_t incr = sizeof(pattern);
>  
> -	incr = sizeof(pattern);
>  	start_phys_aligned = ALIGN(start_phys, incr);
>  	start = __va(start_phys_aligned);
> -	end = (u64 *) (start + size - (start_phys_aligned - start_phys));
> +	end = start + (size - (start_phys_aligned - start_phys)) / incr;
>  	start_bad = 0;
>  	last_bad = 0;
>  


Regards,
Andreas