Re: Oprofile : need to adjust PC by 16 bytes

[Eric Dumazet <dada1@cosmosbay.com>]

Mikael Pettersson a écrit :
> Eric Dumazet writes:
>  > Very strange results on my Intel E5450 CPU
>  > 
>  > I found that I have to bias regs->ip by 16 bytes in order to have
>  > correct profiling.
>  > 
>  > Does it rings a bell to anyone ?
> 
> The term is "skid". All OO processors are susceptible to it,
> as was known already to the Alpha "continous profiling"
> performance counter folks.

Cool. Maybe this can explain the other bug I have with oprofile.

> 
> And no, blindly subtracting 16 from IP is not a fix.

Who mentioned a fix ? I am only giving more fuel to Intel guys so they
hopefully can give us a working oprofile.


> 
>  > 
>  > opcontrol --version
>  > opcontrol: oprofile 0.9.4 compiled on Nov  3 2008 06:45:57
>  > 
>  > Thank you
>  > 
>  > diff --git a/arch/x86/oprofile/op_model_ppro.c b/arch/x86/oprofile/op_model_ppro.c
>  > index 3f1b81a..051fb16 100644
>  > --- a/arch/x86/oprofile/op_model_ppro.c
>  > +++ b/arch/x86/oprofile/op_model_ppro.c
>  > @@ -131,7 +131,9 @@ static int ppro_check_ctrs(struct pt_regs * const regs,
>  >                         continue;
>  >                 rdmsrl(msrs->counters[i].addr, val);
>  >                 if (CTR_OVERFLOWED(val)) {
>  > +                       regs->ip -= 0x10;
>  >                         oprofile_add_sample(regs, i);
>  > +                       regs->ip += 0x10;
>  >                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
>  >                 }
>  >         }
>  > 
>  > --