Re: [Xen-devel] [PATCH v2 3/4] x86/nvmx: split updating RVI from SVI in nvmx_update_apicv

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Mar 26, 2020 at 03:13:56AM +0000, Tian, Kevin wrote:
> > From: Roger Pau Monne <roger.pau@xxxxxxxxxx>
> > Sent: Wednesday, March 25, 2020 6:19 PM
> > 
> > Updating SVI is required when an interrupt has been injected using the
> > Ack on exit VMEXIT feature, so that the in service interrupt in the
> > GUEST_INTR_STATUS matches the vector that is signaled in
> > VM_EXIT_INTR_INFO.
> > 
> > Updating RVI however is not tied to the Ack on exit feature, as it
> > signals the next vector to be injected, and hence should always be
> > updated to the next pending vector, regardless of whether Ack on exit
> > is enabled.
> > 
> > Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> > ---
> >  xen/arch/x86/hvm/vmx/vvmx.c | 22 +++++++++++++++++-----
> >  1 file changed, 17 insertions(+), 5 deletions(-)
> > 
> > diff --git a/xen/arch/x86/hvm/vmx/vvmx.c b/xen/arch/x86/hvm/vmx/vvmx.c
> > index 1753005c91..8431c912a1 100644
> > --- a/xen/arch/x86/hvm/vmx/vvmx.c
> > +++ b/xen/arch/x86/hvm/vmx/vvmx.c
> > @@ -1384,26 +1384,38 @@ static void nvmx_update_apicv(struct vcpu *v)
> >      struct nestedvmx *nvmx = &vcpu_2_nvmx(v);
> >      unsigned long reason = get_vvmcs(v, VM_EXIT_REASON);
> >      unsigned long intr_info = get_vvmcs(v, VM_EXIT_INTR_INFO);
> > +    int rvi;
> > 
> >      if ( reason == EXIT_REASON_EXTERNAL_INTERRUPT &&
> >           nvmx->intr.source == hvm_intsrc_lapic &&
> >           (intr_info & INTR_INFO_VALID_MASK) )
> >      {
> >          uint16_t status;
> > -        uint32_t rvi, ppr;
> > -        uint32_t vector = intr_info & 0xff;
> > +        uint32_t ppr;
> > +        unsigned int vector = intr_info & INTR_INFO_VECTOR_MASK;
> >          struct vlapic *vlapic = vcpu_vlapic(v);
> > 
> > +        /*
> > +         * Update SVI to record the currently in service interrupt that's
> > +         * signaled in EXIT_INTR_INFO.
> > +         */
> >          vlapic_ack_pending_irq(v, vector, 1);
> > 
> >          ppr = vlapic_set_ppr(vlapic);
> >          WARN_ON((ppr & 0xf0) != (vector & 0xf0));
> > 
> >          status = vector << VMX_GUEST_INTR_STATUS_SVI_OFFSET;
> > -        rvi = vlapic_has_pending_irq(v);
> > -        if ( rvi != -1 )
> > -            status |= rvi & VMX_GUEST_INTR_STATUS_SUBFIELD_BITMASK;
> > +        __vmwrite(GUEST_INTR_STATUS, status);
> > +    }
> > +
> > +    rvi = vlapic_has_pending_irq(v);
> > +    if ( rvi != -1 )
> > +    {
> > +        unsigned long status;
> > 
> > +        __vmread(GUEST_INTR_STATUS, &status);
> > +        status &= ~VMX_GUEST_INTR_STATUS_SUBFIELD_BITMASK;
> > +        status |= rvi & VMX_GUEST_INTR_STATUS_SUBFIELD_BITMASK;
> >          __vmwrite(GUEST_INTR_STATUS, status);
> >      }
> 
> I have two minor comments. First, original code requires only one
> __vmwrite but now needs two writes and one read for Ack-on-exit
> is set. Can we optimize to sustain the original behavior?

I think I can manage to do a single write by using a global status
variable initialized to 0 and only do the vmwrite if the value is != 0
by the end of the function.

> Second,
> although I didn't come up a case where always updating RVI (w/
> pending interrupt) may bring some bad effect, it is anyway safer
> to put the whole logic within the check of vmexit reason and intr 
> source.

I also considered this and decided there was no harm in always
updating RVI if there's a pending interrupt, it could even prevent a
future VMEXIT to update RVI AFAICT?

If you prefer I can return early from the function if reason !=
EXIT_REASON_EXTERNAL_INTERRUPT.

Thanks, Roger.