Re: [PATCH v1 05/18] KVM: selftests/hardware_disable_test: code consolidation and cleanup

[Sean Christopherson]

On Thu, Oct 27, 2022, Wang, Wei W wrote:
> On Thursday, October 27, 2022 8:16 AM, Sean Christopherson wrote:
> > > diff --git a/tools/testing/selftests/kvm/hardware_disable_test.c
> > >  static void run_test(uint32_t run)
> > >  {
> > >  	struct kvm_vcpu *vcpu;
> > >  	struct kvm_vm *vm;
> > >  	cpu_set_t cpu_set;
> > > -	pthread_t threads[VCPU_NUM];
> > >  	pthread_t throw_away;
> > > -	void *b;
> > > +	pthread_attr_t attr;
> > >  	uint32_t i, j;
> > > +	int r;
> > >
> > >  	CPU_ZERO(&cpu_set);
> > >  	for (i = 0; i < VCPU_NUM; i++)
> > >  		CPU_SET(i, &cpu_set);
> > 
> > Uh, what is this test doing?  I assume the intent is to avoid spamming all
> > pCPUs in the system, but I don't get the benefit of doing so.
> 
> IIUIC, it is to test if the condition race between the 2 paths:
> #1 kvm_arch_hardware_disable->drop_user_return_notifiers() and
> #2 fire_user_return_notifiers->kvm_on_user_return
> has been solved by disabling interrupts in kvm_on_user_return.
> 
> To stress the tests, it creates a bunch of threads (continuously making syscalls
> to trigger #2 above) to be scheduled on the same pCPU that runs a vCPU, and
> then VM is killed, which triggers #1 above. 
> They fork to test 512 times hoping there is chance #1 and #2 above can happen
> at the same time without an issue.

But why does it matter what pCPU a vCPU is running on?  Wouldn't the probability
of triggering a race between kvm_on_user_return() and hardware_disable() be
_higher_ if there are more pCPUs returning to userspace?