On Tue, Feb 25, 2025 at 07:04:55AM -0800, Sean Christopherson wrote:
> On Tue, Feb 25, 2025, Yan Zhao wrote:
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 58b82d6fd77c..045c61cc7e54 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -12890,11 +12890,11 @@ void kvm_arch_destroy_vm(struct kvm *kvm)
> > > mutex_unlock(&kvm->slots_lock);
> > > }
> > > kvm_unload_vcpu_mmus(kvm);
> > > + kvm_destroy_vcpus(kvm);
> > > kvm_x86_call(vm_destroy)(kvm);
> > > kvm_free_msr_filter(srcu_dereference_check(kvm->arch.msr_filter, &kvm->srcu, 1));
> > > kvm_pic_destroy(kvm);
> > > kvm_ioapic_destroy(kvm);
> > > - kvm_destroy_vcpus(kvm);
> > > kvfree(rcu_dereference_check(kvm->arch.apic_map, 1));
> > > kfree(srcu_dereference_check(kvm->arch.pmu_event_filter, &kvm->srcu, 1));
> > > kvm_mmu_uninit_vm(kvm);
> > After this change, now the sequence is that
> >
> > 1. kvm_arch_pre_destroy_vm()
> > 2. kvm_arch_destroy_vm()
> > 2.1 kvm_destroy_vcpus()
> > 2.2 .vm_destroy hook
> > 2.3 kvm_mmu_uninit_vm() --> mirror root ref is 1 upon here. Zap the mirror
> > root and reclaim SETP page table pages.
> > 2.4 .vm_free hook
> >
> > Since TDX needs to reclaim the TDR page after reclaiming all other pages, we
> > currently added a vm_free hook at 2.4, after 2.3.
> >
> > Could we move kvm_mmu_uninit_vm() before the .vm_destroy hook and after
> > kvm_destroy_vcpus()?
> >
> > Or move the .vm_destroy hook after kvm_mmu_uninit_vm(), e.g. after
> > kvm_page_track_cleanup()?
>
> I would go for the first option. I'll tack on a patch since I need to test all
> of these flows anyways, and I would much prefer to change course sooner rather
> than later if it doesn't work for whatever reason.
>
> Is this comment accurate?
Yes.
And since tdx_mmu_release_hkid() is called before kvm_unload_vcpu_mmus(),
patch 4 in this series is required by TDX. Otherwise, the list_add in
tdx_vcpu_load will cause corruption during tearing down, since
tdx_mmu_release_hkid() has already invoked tdx_disassociate_vp() on all
vcpus.
kvm_unload_vcpu_mmu
vcpu_load
tdx_vcpu_load
list_add(&tdx->cpu_list, &per_cpu(associated_tdvcpus, cpu))
So, maybe a change as below is also required by TDX in case vcpu_load() is
accidentally invoked after .vm_pre_destroy.
diff --git a/arch/x86/kvm/vmx/tdx.c b/arch/x86/kvm/vmx/tdx.c
index b67df0af64f3..183192706ced 100644
--- a/arch/x86/kvm/vmx/tdx.c
+++ b/arch/x86/kvm/vmx/tdx.c
@@ -703,7 +704,7 @@ void tdx_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
struct vcpu_tdx *tdx = to_tdx(vcpu);
vmx_vcpu_pi_load(vcpu, cpu);
- if (vcpu->cpu == cpu)
+ if (vcpu->cpu == cpu || !is_hkid_assigned(to_kvm_tdx(vcpu->kvm)))
return;
tdx_flush_vp_on_cpu(vcpu);
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 1e5f6f820c0b..f5685f153e08 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -12874,13 +12874,19 @@ void kvm_arch_destroy_vm(struct kvm *kvm)
> mutex_unlock(&kvm->slots_lock);
> }
> kvm_destroy_vcpus(kvm);
> +
> + /*
> + * Do final MMU teardown prior to calling into vendor code. All pages
> + * that were donated to the TDX module, e.g. for S-EPT tables, need to
> + * be reclaimed before the VM metadata page can be freed.
> + */
> + kvm_mmu_uninit_vm(kvm);
> kvm_x86_call(vm_destroy)(kvm);
> kvm_free_msr_filter(srcu_dereference_check(kvm->arch.msr_filter, &kvm->srcu, 1));
> kvm_pic_destroy(kvm);
> kvm_ioapic_destroy(kvm);
> kvfree(rcu_dereference_check(kvm->arch.apic_map, 1));
> kfree(srcu_dereference_check(kvm->arch.pmu_event_filter, &kvm->srcu, 1));
> - kvm_mmu_uninit_vm(kvm);
> kvm_page_track_cleanup(kvm);
> kvm_xen_destroy_vm(kvm);
> kvm_hv_destroy_vm(kvm);
