Re: [PATCH] kvm: x86/mmu: Fix is_tdp_mmu_check when using PAE

[Ben Gardon]

On Fri, Nov 13, 2020 at 12:58 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>
> On 11/11/20 19:53, Ben Gardon wrote:
> > When PAE is in use, the root_hpa will not have a shadow page assoicated
> > with it. In this case the kernel will crash with a NULL pointer
> > dereference. Add checks to ensure is_tdp_mmu_root works as intended even
> > when using PAE.
> >
> > Tested: compiles
> >
> > Fixes: 02c00b3a2f7e ("kvm: x86/mmu: Allocate and free TDP MMU roots")
> > Reported-by: Zdenek Kaspar <zkaspar82@xxxxxxxxx>
> > Signed-off-by: Ben Gardon <bgardon@xxxxxxxxxx>
> > ---
> >   arch/x86/kvm/mmu/tdp_mmu.c | 10 ++++++++++
> >   1 file changed, 10 insertions(+)
> >
> > diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> > index 27e381c9da6c..13013f4d98ad 100644
> > --- a/arch/x86/kvm/mmu/tdp_mmu.c
> > +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> > @@ -49,8 +49,18 @@ bool is_tdp_mmu_root(struct kvm *kvm, hpa_t hpa)
> >   {
> >       struct kvm_mmu_page *sp;
> >
> > +     if (WARN_ON(!VALID_PAGE(hpa)))
> > +             return false;
> > +
> >       sp = to_shadow_page(hpa);
> >
> > +     /*
> > +      * If this VM is being run with PAE, the TDP MMU will not be enabled
> > +      * and the root HPA will not have a shadow page associated with it.
> > +      */
> > +     if (!sp)
> > +             return false;
> > +
> >       return sp->tdp_mmu_page && sp->root_count;
> >   }
> >
> >
>
> If this was just PAE, it would be easier to test "if (shadow_root_level
>  >= PT64_ROOT_4LEVEL)"---and more correct too, because using the
> page_private of __pa(vcpu->arch.mmu->pae_root) is a bit untidy; we
> should only use page_private for pages that we know have a shadow page.
>
> In Jamie's case however, it is x86_64 (so kvm_mmu_get_tdp_level(vcpu) ==
> 4 and therefore the "if (shadow_root_level >= PT64_ROOT_4LEVEL)" would
> be true) but without EPT.  In that case we go through
>
>         vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->lm_root);
>
> but lm_root is allocated with get_zeroed_page and therefore
> to_shadow_page is NULL.
>
> I am thinking of testing simply "if (tdp_enabled)" so that we can see if
> there are other cases with to_shadow_page(hpa) == NULL and we don't
> sweep them under the rug.  Or test "if (tdp_enabled)" and also WARN if
> !sp.  What do you think?

That sounds good to me. I see you already mailed out a modified
version of this change, so I'll go review it.
Thanks!

>
> Paolo
>