Re: linux-next: Tree for Aug 26 (i7core_edac)

[Borislav Petkov <bp@amd64.org>]

On Fri, Aug 26, 2011 at 10:16:11AM -0400, Sedat Dilek wrote:
> ( UNTESTED, just followed patches like "[PATCH] p54: Use do_div for
> 64-bit division to fix 32-bit kernels" or "[PATCH v1] carl9170: Use
> do_div for 64-bit division to fix 32-bit kernels" )
> 
> Regards,
> - Sedat -

> diff --git a/drivers/edac/i7core_edac.c b/drivers/edac/i7core_edac.c
> index 7cb68de..4d4f3a5 100644
> --- a/drivers/edac/i7core_edac.c
> +++ b/drivers/edac/i7core_edac.c
> @@ -37,6 +37,7 @@
>  #include <linux/smp.h>
>  #include <asm/mce.h>
>  #include <asm/processor.h>
> +#include <asm/div64.h>
>  
>  #include "edac_core.h"
>  
> @@ -2102,7 +2103,8 @@ static int set_sdram_scrub_rate(struct mem_ctl_info *mci, u32 new_bw)
>  		 * program the corresponding register value.
>  		 */
>  		scrub_interval = (unsigned long long)freq_dclk_mhz *
> -			cache_line_size * 1000000 / new_bw;
> +			cache_line_size * 1000000);
> +		do_div(scrub_interval, new_bw);

This has been an issue pretty often lately on lkml. Instead of doing
64-bit division on 32-bit (which is really slow), you could try to use
the commutative property of integer multiplication and thus reorder the
operations so that the division is 32-bit only. Maybe

((freq_dclk_mhz * cache_line_size) / new_bw) * 1000000

for example. You have to make sure though that new_bw as divisor can
never be bigger than the dividend because otherwise you get the 0. One
question would be what is the interval of freq_dclk_mhz and can (new_bw >> 6)
be ever bigger than it.

And while at it, the cache_line_size is pretty solidly staying 64 Byte
in the short-term future so no need for a local variable. Also, shifting
left by 6 should be a faster equivalent than multiplying by 64.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551