Re: [PATCH 17/15] KVM: X86: Ensure pae_root to be reconstructed for shadow paging if the guest PDPTEs is changed

[Sean Christopherson <seanjc@google.com>]

On Thu, Dec 09, 2021, Paolo Bonzini wrote:
> On 12/8/21 01:15, Sean Christopherson wrote:
> > > @@ -832,8 +832,14 @@ int load_pdptrs(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, unsigned long cr3)
> > >   	if (memcmp(mmu->pdptrs, pdpte, sizeof(mmu->pdptrs))) {
> > >   		memcpy(mmu->pdptrs, pdpte, sizeof(mmu->pdptrs));
> > >   		kvm_register_mark_dirty(vcpu, VCPU_EXREG_PDPTR);
> > > -		/* Ensure the dirty PDPTEs to be loaded. */
> > > -		kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
> > > +		/*
> > > +		 * Ensure the dirty PDPTEs to be loaded for VMX with EPT
> > > +		 * enabled or pae_root to be reconstructed for shadow paging.
> > > +		 */
> > > +		if (tdp_enabled)
> > > +			kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
> > > +		else
> > > +			kvm_mmu_free_roots(vcpu, vcpu->arch.mmu, KVM_MMU_ROOT_CURRENT);
> > Shouldn't matter since it's legacy shadow paging, but @mmu should be used instead
> > of vcpu->arch.mmuvcpu->arch.mmu.
> 
> In kvm/next actually there's no mmu parameter to load_pdptrs, so it's okay
> to keep vcpu->arch.mmu.
> 
> > To avoid a dependency on the previous patch, I think it makes sense to have this be:
> > 
> > 	if (!tdp_enabled && memcmp(mmu->pdptrs, pdpte, sizeof(mmu->pdptrs)))
> > 		kvm_mmu_free_roots(vcpu, mmu, KVM_MMU_ROOT_CURRENT);
> > 
> > before the memcpy().
> > 
> > Then we can decide independently if skipping the KVM_REQ_LOAD_MMU_PGD if the
> > PDPTRs are unchanged with respect to the MMU is safe.
> 
> Do you disagree that there's already an invariant that the PDPTRs can only
> be dirty if KVM_REQ_LOAD_MMU_PGD---and therefore a previous change to the
> PDPTRs would have triggered KVM_REQ_LOAD_MMU_PGD?

What I think is moot, because commit 24cd19a28cb7 ("KVM: X86: Update mmu->pdptrs
only when it is changed") breaks nested VMs with EPT in L0 and PAE shadow paging
in L2.  Reproducing is trivial, just disable EPT in L1 and run a VM.  I haven't
investigating how it breaks things, because why it's broken is secondary for me.

My primary concern is that we would even consider optimizing the PDPTR logic without
a mountain of evidence that any patch is correct for all scenarios.  We had to add
an entire ioctl() just to get PDPTRs functional.  This apparently wasn't validated
against a simple use case, let alone against things like migration with nested VMs,
multliple L2s, etc...