Re: tip.today - scheduler bam boom crash (cpu hotplug)

[Peter Zijlstra]

On Mon, Feb 27, 2017 at 02:06:27PM -0500, Paolo Bonzini wrote:
> > > +	if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
> > > +		mark_tsc_unstable("not invariant");
> > 
> > Errm, no.
> > 
> > That makes TSC unusable for systems which do not go into C/P states in
> > which the TSC stops. There is a world outside KVM ....
> 
> Actually I was surprised too by Peter's patch, as it was adding
> mark_tsc_unstable pretty much everywhere that didn't have nonstop TSC.
> But hopefully it would still be okay to call clear_sched_clock_stable
> in tsc_init, in the same way.

I was preserving the sched_clock behaviour. That said, maybe we can
simply remove all those extra checks and only rely on the current
mark_tsc_unstable() calls.

I'll just have to sit down and consider the various cases..