Re: [PATCH 3/9] KVM: selftests: Affine child tasks to other pCPUs in hardware disable test

On Thu, Jul 30, 2026, [email protected] wrote:
> > diff --git a/tools/testing/selftests/kvm/hardware_disable_test.c b/tools/testing/selftests/kvm/hardware_disable_test.c
> > index 3147f5c97e943..1bcbd13e99947 100644
> > --- a/tools/testing/selftests/kvm/hardware_disable_test.c
> > +++ b/tools/testing/selftests/kvm/hardware_disable_test.c
> [ ... ]
> > @@ -84,16 +85,11 @@ static void run_test(u32 run)
> [ ... ]
> >  		check_create_thread(&threads[i], NULL, run_vcpu, vcpu);
> > -		check_set_affinity(threads[i], &cpu_set);
> > +		check_set_affinity(threads[i], &child_cpu_set);
> >  
> >  		for (j = 0; j < SLEEPING_THREAD_NUM; ++j) {
> >  			check_create_thread(&throw_away, NULL, sleeping_thread,
> >  					    (void *)NULL);
> > -			check_set_affinity(throw_away, &cpu_set);
> > +			check_set_affinity(throw_away, &child_cpu_set);
> >  		}
> >  	}
> [ ... ]
> > @@ -147,12 +143,30 @@ void wait_for_child_setup(pid_t pid)
> [ ... ]
> > +static void setup_child_cpu_set(void)
> > +{
> > +	int cpu;
> > +
> > +	kvm_sched_getaffinity(0, sizeof(child_cpu_set), &child_cpu_set);
> > +
> > +	if (CPU_COUNT(&child_cpu_set) < 2)
> > +		return;
> > +
> > +	cpu = pin_task_to_random_cpu(pthread_self(), &child_cpu_set);
> > +	CPU_CLR(cpu, &child_cpu_set);
> > +
> > +	while (CPU_COUNT(&child_cpu_set) > VCPU_NUM)
> > +		CPU_CLR(kvm_pick_random_cpu(&child_cpu_set), &child_cpu_set);
> > +}
> 
> [Severity: Medium]
> Since setup_child_cpu_set() pins the parent thread to a single random CPU
> prior to spawning children, won't the newly created threads in run_test()
> inherit this single-CPU affinity upon creation?
> 
> Because check_create_thread() spawns the threads before their affinity is
> explicitly updated to child_cpu_set, they might immediately begin executing
> on the exact same CPU as the parent.

Yeah, I was trying to avoid this since pthread_attr_setaffinity_np() is only
available with __USE_GNU.

> Could this result in the child threads starving the parent on that single CPU
> before the parent gets a chance to execute check_set_affinity()? This seems
> like it might heavily delay the migration and trigger the same CPU contention
> that this patch aims to fix.

After far too much testing, the problem isn't direct CPU contention, AFAICT it's
due to incurring cross-node NUMA penalties, and presumably scheduler contention
when the pool of possible CPUs to run on is huge.

I tried a bunch of fancy approaches, e.g. making parent, child, and grandchildren
all mutually exclusive, and while the fanciness can provide marginal benefits when
pthread_attr_setaffinity_np() isn't available, all that really matters is that the
worker threads get affined before they start doing work.