Re: [RFC PATCH v1 1/2] sched: unified sched_powersavings sysfs tunable

[Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>]

* Peter Zijlstra <a.p.zijlstra@chello.nl> [2012-01-25 15:53:01]:

> On Mon, 2012-01-16 at 21:52 +0530, Vaidyanathan Srinivasan wrote:
> > +++ b/block/blk.h
> > @@ -167,14 +167,15 @@ static inline int queue_congestion_off_threshold(struct request_queue *q)
> >  static inline int blk_cpu_to_group(int cpu)
> >  {
> >         int group = NR_CPUS;
> > -#ifdef CONFIG_SCHED_MC
> > -       const struct cpumask *mask = cpu_coregroup_mask(cpu);
> > -       group = cpumask_first(mask);
> > -#elif defined(CONFIG_SCHED_SMT)
> > -       group = cpumask_first(topology_thread_cpumask(cpu));
> > +#ifdef CONFIG_SCHED_POWERSAVE
> > +       if (smt_capable())
> > +               group = cpumask_first(topology_thread_cpumask(cpu));
> > +       else    
> > +               group = cpumask_first(cpu_coregroup_mask(cpu));
> >  #else
> >         return cpu;
> >  #endif
> > +       /* Possible dead code?? */
> >         if (likely(group < NR_CPUS))
> >                 return group;
> >         return cpu; 
> 
> After going, WTF is block doing! I had a closer look and this doesn't
> seem right at all. The old code would use coregroup_mask when SCHED_MC
> && SCHED_SMT, the new code does something else.
> 
> Jens, what is this thing trying to do?

I understood the requirement as get first cpu in the 'core' on
a hyper-threaded system and first cpu in the 'socket' on
a non-threaded system for best cache affinity.  Based on Jens
explanation and Peter's patch, identifying last-level shared cache
works best for this case.

--Vaidy