Re: [PATCH V3 01/12] KVM: X86/MMU: Verify PDPTE for nested NPT in PAE paging mode when page fault

[Sean Christopherson]

On Sat, May 21, 2022, Lai Jiangshan wrote:
> From: Lai Jiangshan <jiangshan.ljs@xxxxxxxxxxxx>
> Fixes: e4e517b4be01 ("KVM: MMU: Do not unconditionally read PDPTE from guest memory")
> Signed-off-by: Lai Jiangshan <jiangshan.ljs@xxxxxxxxxxxx>
> ---
>  arch/x86/kvm/mmu/paging_tmpl.h | 39 ++++++++++++++++++++++++++++++++++
>  1 file changed, 39 insertions(+)
> 
> diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h
> index db80f7ccaa4e..6e3df84e8455 100644
> --- a/arch/x86/kvm/mmu/paging_tmpl.h
> +++ b/arch/x86/kvm/mmu/paging_tmpl.h
> @@ -870,6 +870,44 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault
>  	if (is_page_fault_stale(vcpu, fault, mmu_seq))
>  		goto out_unlock;
>  
> +	/*
> +	 * When nested NPT enabled and L1 is PAE paging, mmu->get_pdptrs()
> +	 * which is nested_svm_get_tdp_pdptr() reads the guest NPT's PDPTE
> +	 * from memory unconditionally for each call.
> +	 *
> +	 * The guest PAE root page is not write-protected.

I think it's worth calling out that it's simply not feasible to write-protect
PDPTEs due to them not covering a full page.

And looking at this comment as a whole, while I love detailed comments, I think
it'd be better off to avoid referring to mmu->get_pdptrs() and use more generic
terminology when talking about KVM.

I think this is accurate?

	/*
	 * If KVM is shadowing nested NPT and L1 is using PAE paging, zap the
	 * root for the PDPTE if the cached value doesn't match the entry at the
	 * time of the page fault, and resume the guest to rebuid the root.
	 * This is effectively a variation of write-protection, where the target
	 * SPTE(s) is zapped on use instead of on write.
	 *
	 * Under SVM with NPT+PAE, the CPU does NOT cache PDPTEs and instead
	 * handles them as it would any other page table entry.  I.e. KVM can't
	 * cache PDPTEs at nested VMRUN without violating the SVM architecture.
	 *
	 * KVM doesn't write-protect PDPTEs because CR3 only needs to be 32-byte
	 * aligned and sized when using PAE paging, whereas write-protection
	 * works at page granularity.
	 */

> +	 *
> +	 * The mmu->get_pdptrs() in FNAME(walk_addr_generic) might get a value
> +	 * different from previous calls or different from the return value of
> +	 * mmu->get_pdptrs() in mmu_alloc_shadow_roots().
> +	 *
> +	 * It will cause FNAME(fetch) installs the spte in a wrong sp or links
> +	 * a sp to a wrong parent if the return value of mmu->get_pdptrs()
> +	 * is not verified unchanged since FNAME(gpte_changed) can't check
> +	 * this kind of change.
> +	 *
> +	 * Verify the return value of mmu->get_pdptrs() (only the gfn in it
> +	 * needs to be checked) and do kvm_mmu_free_roots() like load_pdptr()
> +	 * if the gfn isn't matched.
> +	 *
> +	 * Do the verifying unconditionally when the guest is PAE paging no
> +	 * matter whether it is nested NPT or not to avoid complicated code.

Doing this unconditionally just trades one architecturally incorrect behavior
with another.  Does any real world use case actually care?  Probably not.  But the
behavior is visible to software, and I don't think it costs us much to get it right.

There are a number of ways to handle this, e.g. set a flag in kvm_init_shadow_npt_mmu()
and consume it here.  We could probably even burn a bit in kvm_mmu_extended_role
since we have lots of bits to burn.  E.g.

	if (vcpu->arch.mmu->cpu_role.ext.npt_pae) {

	}


> +	 */
> +	if (vcpu->arch.mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) {
> +		u64 pdpte = vcpu->arch.mmu->pae_root[(fault->addr >> 30) & 3];
> +		struct kvm_mmu_page *sp = NULL;
> +
> +		if (IS_VALID_PAE_ROOT(pdpte))
> +			sp = to_shadow_page(pdpte & PT64_BASE_ADDR_MASK);
> +
> +		if (!sp || walker.table_gfn[PT32E_ROOT_LEVEL - 2] != sp->gfn) {
> +			write_unlock(&vcpu->kvm->mmu_lock);
> +			kvm_mmu_free_roots(vcpu->kvm, vcpu->arch.mmu,
> +					   KVM_MMU_ROOT_CURRENT);
> +			goto release_clean;
> +		}
> +	}
> +
>  	r = make_mmu_pages_available(vcpu);
>  	if (r)
>  		goto out_unlock;
> @@ -877,6 +915,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault
>  
>  out_unlock:
>  	write_unlock(&vcpu->kvm->mmu_lock);
> +release_clean:
>  	kvm_release_pfn_clean(fault->pfn);
>  	return r;
>  }
> -- 
> 2.19.1.6.gb485710b
>