Re: [rfc][patch] select_idle_sibling() inducing bouncing on westmere

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

On Sat, May 26, 2012 at 09:29:24AM +0200, Peter Zijlstra wrote:
> On Sat, 2012-05-26 at 08:37 +0200, Mike Galbraith wrote:
> 
> > Ew. 3.4 went broke for Q6600, and performance went... far far away.
> > 
> > [    0.200057] CPU0 attaching sched-domain:
> > [    0.204016]  domain 0: span 0-3 level MC
> > [    0.208015]   groups: 0 1 2 3
> > [    0.210970] CPU1 attaching sched-domain:
> > [    0.212014]  domain 0: span 0-3 level MC
> > [    0.216016]   groups: 1 2 3 0
> > [    0.220016] CPU2 attaching sched-domain:
> > [    0.224015]  domain 0: span 0-3 level MC
> > [    0.228016]   groups: 2 3 0 1
> > [    0.232015] CPU3 attaching sched-domain:
> > [    0.236016]  domain 0: span 0-3 level MC
> > [    0.240017]   groups: 3 0 1 2
> 
> 
> Oh yikes, I guess I wrecked
> arch/x86/kernel/smpboot.c:cpu_coregroup_mask() in
> 8e7fbcbc22c12414bcc9dfdd683637f58fb32759.
> 
> That should very much always return llc mask, I just got that AMD case
> confused. It looks like it should look like:
> 
> 
> const struct cpumask *cpu_coregroup_mask(int cpu)
> {
> 	return cpu_llc_mask(cpu);
> }
> 
> And the AMD_DCM check was just to undo powersavings damage on
> Magny-Cours or somesuch.

IIRC returning cpu_core_mask() could even cause a panic in the
scheduler, because the hierarchy of scheduling groups/domains was
broken.
 
> Andreas?

Returning cpu_llc_mask is the right thing to do on AMD.



Andreas