Re: [PATCH 3/3] KVM: Avoid synchronize_srcu() in kvm_io_bus_register_dev()

[Sean Christopherson <seanjc@google.com>]

On Thu, Jul 10, 2025, Keir Fraser wrote:
> On Mon, Jul 07, 2025 at 11:49:34AM -0700, Sean Christopherson wrote:
> Would it be satisfactory to put a patch along the lines of your
> suggestions below into a v2 of this patch series?

Ya, works for me.

> I have made some comments below.
> 
> > 
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 4953846cb30d..057fb4ce66b0 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -5861,6 +5861,9 @@ static int handle_invalid_guest_state(struct kvm_vcpu *vcpu)
> >                 if (kvm_test_request(KVM_REQ_EVENT, vcpu))
> >                         return 1;
> >  
> > +               /* Or maybe smp_mb()?  Not sure what this needs to be. */
> > +               barrier();
> > +
> 
> Looks weak but maybe strong enough for x86? Maybe smp_rmb() would be better
> statement of intention?

Hmm, yeah, smp_rmb() is better.  I was thinking it just needs to be a compiler
barrier, to ensure KVM reads kvm->buses as needed for each emulated instruction.
But ignoring that x86 is strongly ordered, KVM also needs to ensure a store to
kvm->buses that is supposed to be observed by the next guest instruction is fully
visibile before that instruction executes.

> 
> >                 if (!kvm_emulate_instruction(vcpu, 0))
> >                         return 0;
> >  

...

> I guess kvm_io_bus_read() is to be done as well? Perhaps the barrier
> and dereference should be pulled into a helper with the comment, just
> in one place?

Ya, +1 to a helper.