On Mon, Sep 13, 2021, Maxim Levitsky wrote:
> On Thu, 2021-09-02 at 23:40 +0000, Sean Christopherson wrote:
> > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > index 4853c033e6ce..03293cd3c7ae 100644
> > --- a/arch/x86/kvm/mmu/mmu.c
> > +++ b/arch/x86/kvm/mmu/mmu.c
> > @@ -2143,8 +2143,10 @@ static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
> > kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
> > }
> >
> > - if (sp->unsync_children)
> > - kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
> > + if (sp->unsync_children) {
> > + kvm_make_all_cpus_request(KVM_REQ_MMU_SYNC, vcpu);
>
> I don't know the KVM mmu well so I miss something here most likely,
> but why to switch to kvm_make_all_cpus_request?
>
> MMU is shared by all VCPUs, and the process of its syncing should also do
> remote TLB flushes when needed?
>
> Another thing I don't fully understand is why this patch is needed. If we
> link an SP which has unsync children, we raise KVM_REQ_MMU_SYNC, which I
> think means that *this* vCPU will sync the whole MMU on next guest entry,
> including these unsync child SPs. Could you explain this?
Answering all three questions at once, the problem is that KVM links in a new SP
that points at unsync'd SPs _before_ servicing KVM_REQ_MMU_SYNC. While the vCPU
is guaranteed to service KVM_REQ_MMU_SYNC before entering the guest, that doesn't
hold true for other vCPUs. As a result, there's a window where a different vCPU
can consume the stale, unsync SP via the new SP.
