Re: [PATCH v2 02/11] KVM: nSVM: clean up the copying of V_INTR bits from vmcb02 to vmcb12

[Maxim Levitsky]

On Tue, 2023-01-31 at 01:44 +0000, Sean Christopherson wrote:
> On Sat, Jan 28, 2023, Sean Christopherson wrote:
> > On Tue, Nov 29, 2022, Maxim Levitsky wrote:
> > > the V_IRQ and v_TPR bits don't exist when virtual interrupt
> > > masking is not enabled, therefore the KVM should not copy these
> > > bits regardless of V_IRQ intercept.
> > 
> > Hmm, the APM disagrees:

Yes, my apologies, after re-reading the APM I agree with you.


> > 
> >  The APIC's TPR always controls the task priority for physical interrupts, and the
> >  V_TPR always controls virtual interrupts.
> > 
> >    While running a guest with V_INTR_MASKING cleared to 0:
> >      • Writes to CR8 affect both the APIC's TPR and the V_TPR register.
> > 
> > 
> >  ...
> > 
> >  The three VMCB fields V_IRQ, V_INTR_PRIO, and V_INTR_VECTOR indicate whether there
> >  is a virtual interrupt pending, and, if so, what its vector number and priority are.
> > 
> > IIUC, V_INTR_MASKING_MASK is mostly about EFLAGS.IF, with a small side effect on
> > TPR.  E.g. a VMM could pend a V_IRQ but clear V_INTR_MASKING and expect the guest
> > to take the V_IRQ.  At least, that's my reading of things.

Yes, this is how I understand it as well.


> > 
> > > Signed-off-by: Maxim Levitsky <mlevitsk@xxxxxxxxxx>
> > > ---
> > >  arch/x86/kvm/svm/nested.c | 23 ++++++++---------------
> > >  1 file changed, 8 insertions(+), 15 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> > > index 37af0338da7c32..aad3145b2f62fe 100644
> > > --- a/arch/x86/kvm/svm/nested.c
> > > +++ b/arch/x86/kvm/svm/nested.c
> > > @@ -412,24 +412,17 @@ void nested_copy_vmcb_save_to_cache(struct vcpu_svm *svm,
> > >   */
> > >  void nested_sync_control_from_vmcb02(struct vcpu_svm *svm)
> > >  {
> > > -	u32 mask;
> > > +	u32 mask = 0;
> > >  	svm->nested.ctl.event_inj      = svm->vmcb->control.event_inj;
> > >  	svm->nested.ctl.event_inj_err  = svm->vmcb->control.event_inj_err;
> > >  
> > > -	/* Only a few fields of int_ctl are written by the processor.  */
> > > -	mask = V_IRQ_MASK | V_TPR_MASK;
> > > -	if (!(svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK) &&
> > > -	    svm_is_intercept(svm, INTERCEPT_VINTR)) {
> > > -		/*
> > > -		 * In order to request an interrupt window, L0 is usurping
> > > -		 * svm->vmcb->control.int_ctl and possibly setting V_IRQ
> > > -		 * even if it was clear in L1's VMCB.  Restoring it would be
> > > -		 * wrong.  However, in this case V_IRQ will remain true until
> > > -		 * interrupt_window_interception calls svm_clear_vintr and
> > > -		 * restores int_ctl.  We can just leave it aside.
> > > -		 */
> > > -		mask &= ~V_IRQ_MASK;
> 
> Argh! *shakes fist at KVM and SVM*
> 
> This is ridiculously convoluted, and I'm pretty sure there are existing bugs.  If
> L1 runs L2 with V_IRQ=1 and V_INTR_MASKING=1


Note that there are two cases when we need an interrupt window in nested case:

- If the L1 doesn't intercept interrupts, which is what we are taking about here.
- If the L1 does intercept interrupts, but let L2 control the L1's EFLAGS.IF and/or
  L1's GIF 
  (that is V_INTR_MASKING_MASK is not set, and/or L1 doesn't intercept STGI/CLGI).

  In this case a 'real' interrupt will be converted to a VM exit but only
  when both L1's EFLAGS.IF is true and L1's GIF is true.



> , and KVM requests an interrupt window,
> then KVM will overwrite vmcb02's int_vector and int_ctl, i.e. clobber L1's V_IRQ,
> but then silently clear INTERCEPT_VINTR in recalc_intercepts() and thus prevent
> svm_clear_vintr() from being reached, i.e. prevent restoring L1's V_IRQ.


> 
> Bug #1 is that KVM shouldn't clobber the V_IRQ fields if KVM ultimately decides
> not to open an interrupt window.  Bug #2 is that KVM needs to open an interrupt
> window if save.RFLAGS.IF=1, as interrupts may become unblocked in that case,
> e.g. if L2 is in an interrupt shadow.


> 
> So I think this over two patches?
> 
> diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> index 05d38944a6c0..ad1e70ac8669 100644
> --- a/arch/x86/kvm/svm/nested.c
> +++ b/arch/x86/kvm/svm/nested.c
> @@ -139,13 +139,18 @@ void recalc_intercepts(struct vcpu_svm *svm)
>  
>         if (g->int_ctl & V_INTR_MASKING_MASK) {
>                 /*
> -                * Once running L2 with HF_VINTR_MASK, EFLAGS.IF and CR8
> -                * does not affect any interrupt we may want to inject;
> -                * therefore, writes to CR8 are irrelevant to L0, as are
> -                * interrupt window vmexits.
> +                * If L2 is active and V_INTR_MASKING is enabled in vmcb12,
> +                * disable intercept of CR8 writes as L2's CR8 does not affect
> +                * any interrupt KVM may want to inject.
> +                *
> +                * Similarly, disable intercept of virtual interrupts (used to
> +                * detect interrupt windows) if the saved RFLAGS.IF is '0', as
> +                * the effective RFLAGS.IF for L1 interrupts will never be set
> +                * while L2 is running (L2's RFLAGS.IF doesn't affect L1 IRQs).
>                  */
>                 vmcb_clr_intercept(c, INTERCEPT_CR8_WRITE);
> -               vmcb_clr_intercept(c, INTERCEPT_VINTR);
> +               if (!(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF))
> +                       vmcb_clr_intercept(c, INTERCEPT_VINTR);

How about instead moving this code to svm_set_vintr?

That is, in the guest mode, if the guest has V_INTR_MASKING_MASK, then
then a nested VM exit is the next point the interrupt window could open,
thus we don't set VINTR)

Or even better put the logic in svm_enable_irq_window (that is avoid
calling svm_set_vintr in the first place).

I also think that it worth it to add a warning that 'svm_set_intercept'
didn't work, that is didn't really set an intercept. 
In theory that can result in nasty CVEs in addition to logic bugs as you found.


>         }
>  
>         /*
> @@ -416,18 +421,18 @@ void nested_sync_control_from_vmcb02(struct vcpu_svm *svm)
>  
>         /* Only a few fields of int_ctl are written by the processor.  */
>         mask = V_IRQ_MASK | V_TPR_MASK;
> -       if (!(svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK) &&
> -           svm_is_intercept(svm, INTERCEPT_VINTR)) {
> -               /*
> -                * In order to request an interrupt window, L0 is usurping
> -                * svm->vmcb->control.int_ctl and possibly setting V_IRQ
> -                * even if it was clear in L1's VMCB.  Restoring it would be
> -                * wrong.  However, in this case V_IRQ will remain true until
> -                * interrupt_window_interception calls svm_clear_vintr and
> -                * restores int_ctl.  We can just leave it aside.
> -                */
> +
> +       /*
> +        * Don't sync vmcb02 V_IRQ back to vmcb12 if KVM (L0) is intercepting
> +        * virtual interrupts in order to request an interrupt window, as KVM
> +        * has usurped vmcb02's int_ctl.  If an interrupt window opens before
> +        * the next VM-Exit, svm_clear_vintr() will restore vmcb12's int_ctl.
> +        * If no window opens, V_IRQ will be correctly preserved in vmcb12's
> +        * int_ctl (because it was never recognized while L2 was running).
> +        */
> +       if (svm_is_intercept(svm, INTERCEPT_VINTR) &&
> +           !test_bit(INTERCEPT_VINTR, (unsigned long *)svm->nested.ctl.intercepts))
>                 mask &= ~V_IRQ_MASK;

This makes sense.



> -       }
>  
>         if (nested_vgif_enabled(svm))
>                 mask |= V_GIF_MASK;
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index b103fe7cbc82..59d2891662ef 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -1580,6 +1580,16 @@ static void svm_set_vintr(struct vcpu_svm *svm)
>  
>         svm_set_intercept(svm, INTERCEPT_VINTR);
>  
> +       /*
> +        * Recalculating intercepts may have clear the VINTR intercept.  If
> +        * V_INTR_MASKING is enabled in vmcb12, then the effective RFLAGS.IF
> +        * for L1 physical interrupts is L1's RFLAGS.IF at the time of VMRUN.
> +        * Requesting an interrupt window if save.RFLAGS.IF=0 is pointless as
> +        * interrupts will never be unblocked while L2 is running.
> +        */
> +       if (!svm_is_intercept(svm, INTERCEPT_VINTR))
> +               return;

This won't be needed if we don't call the svm_set_vintr in the first place.

> +
>         /*
>          * This is just a dummy VINTR to actually cause a vmexit to happen.
>          * Actual injection of virtual interrupts happens through EVENTINJ.
> 



With all this said, I also want to note that this patch has *nothing* to do with VNMI,
I only added it due to some refactoring, so feel free to drop it from vNMI queue,
and deal with those bugs separately.

Best regards,
	Maxim Levitsky