2004-09-19 14:07:12

by Dominik Brodowski

Subject: boot_cpu_data vs current_cpu_data in voluntary-preempt-2.6.9-rc2-mm1-S1

Ingo,

Your voluntary-preempt-2.6.9-rc2-mm1-S1 patch contains this change

@@ -34,7 +34,7 @@ inline void __const_udelay(unsigned long
xloops *= 4;
__asm__("mull %0"
:"=d" (xloops), "=&a" (d0)
- :"1" (xloops),"0" (current_cpu_data.loops_per_jiffy *
(HZ/4)));
+ :"1" (xloops),"0" (boot_cpu_data.loops_per_jiffy * (HZ/4)));
__delay(++xloops);
}

for both x86 and x86_64. And it's wrong. It assumes loops_per_jiffy being
consistent on all CPUs. There _are_ asymetric multiprocessor systems out
there, and some SMP systems can become asymetric as soon as
frequency scaling is enabled. Using boot_cpu_data's loops_per_jiffy instead
of current_cpu_data's loops_per_jiffy causes delays being too short or too
long. So please drop this change.

Thanks,
Dominik

2004-09-19 17:46:01

by Ingo Molnar

Subject: Re: boot_cpu_data vs current_cpu_data in voluntary-preempt-2.6.9-rc2-mm1-S1

On Sun, 19 Sep 2004, Dominik Brodowski wrote:

> Your voluntary-preempt-2.6.9-rc2-mm1-S1 patch contains this change
>
> @@ -34,7 +34,7 @@ inline void __const_udelay(unsigned long
> xloops *= 4;
> __asm__("mull %0"
> :"=d" (xloops), "=&a" (d0)
> - :"1" (xloops),"0" (current_cpu_data.loops_per_jiffy *
> (HZ/4)));
> + :"1" (xloops),"0" (boot_cpu_data.loops_per_jiffy * (HZ/4)));

this comes from the BKL patch - this is done to avoid false positives in
the smp_processor_id() debugger.

Ingo