Re: [PATCH v2 5/6] KVM-GST: adjust scheduler cpu power

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2011-01-31 at 12:25 +0100, Peter Zijlstra wrote:
> On Fri, 2011-01-28 at 14:52 -0500, Glauber Costa wrote:
> 
> > +#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
> > +static DEFINE_PER_CPU(u64, cpu_steal_time);
> > +
> > +#ifndef CONFIG_64BIT
> > +static DEFINE_PER_CPU(seqcount_t, steal_time_seq);
> > +
> > +static inline void steal_time_write_begin(void)
> > +{
> > +	__this_cpu_inc(steal_time_seq.sequence);
> > +	smp_wmb();
> > +}
> > +
> > +static inline void steal_time_write_end(void)
> > +{
> > +	smp_wmb();
> > +	__this_cpu_inc(steal_time_seq.sequence);
> > +}
> > +
> > +static inline u64 steal_time_read(int cpu)
> > +{
> > +	u64 steal_time;
> > +	unsigned seq;
> > +
> > +	do {
> > +		seq = read_seqcount_begin(&per_cpu(steal_time_seq, cpu));
> > +		steal_time = per_cpu(cpu_steal_time, cpu);
> > +	} while (read_seqcount_retry(&per_cpu(steal_time_seq, cpu), seq));
> > +
> > +	return steal_time;
> > +}
> > +#else /* CONFIG_64BIT */
> > +static inline void steal_time_write_begin(void)
> > +{
> > +}
> > +
> > +static inline void steal_time_write_end(void)
> > +{
> > +}
> > +
> > +static inline u64 steal_time_read(int cpu)
> > +{
> > +	return per_cpu(cpu_steal_time, cpu);
> > +}
> 
> 
> > @@ -3536,6 +3592,11 @@ static int touch_steal_time(int is_idle)
> >  
> >  	if (st) {
> >  		account_steal_time(st);
> > +#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
> > +		steal_time_write_begin();
> > +		__this_cpu_add(cpu_steal_time, steal);
> > +		steal_time_write_end();
> > +#endif
> >  		return 1;
> >  	}
> >  	return 0;
> 
> 
> Why replicate all logic you've already got in patch 4? That too is
> reading steal time in a loop in kvm_account_steal_time(), why not extend
> that interface to take a cpu argument and be done with it?

Also, this relies on touch_steal_time() to have ran, not at all
something that is tied to calling update_rq_clock().

We can call update_rq_clock() many many times between ticks, as can the
stealtime increase because the vcpu can schedule many many times between
ticks.