Re: [PATCH v1 1/3] KVM: VMX: Consider PID.PIR to determine if vCPU has pending interrupts

[Paolo Bonzini]

On 06/11/19 18:56, Joao Martins wrote:
> Commit 17e433b54393 ("KVM: Fix leak vCPU's VMCS value into other pCPU")
> introduced vmx_dy_apicv_has_pending_interrupt() in order to determine
> if a vCPU have a pending posted interrupt. This routine is used by
> kvm_vcpu_on_spin() when searching for a a new runnable vCPU to schedule
> on pCPU instead of a vCPU doing busy loop.
> 
> vmx_dy_apicv_has_pending_interrupt() determines if a
> vCPU has a pending posted interrupt solely based on PID.ON. However,
> when a vCPU is preempted, vmx_vcpu_pi_put() sets PID.SN which cause
> raised posted interrupts to only set bit in PID.PIR without setting
> PID.ON (and without sending notification vector), as depicted in VT-d
> manual section 5.2.3 "Interrupt-Posting Hardware Operation".
> 
> Therefore, checking PID.ON is insufficient to determine if a vCPU has
> pending posted interrupts and instead we should also check if there is
> some bit set on PID.PIR.
> 
> Fixes: 17e433b54393 ("KVM: Fix leak vCPU's VMCS value into other pCPU")
> Signed-off-by: Joao Martins <joao.m.martins@xxxxxxxxxx>
> Signed-off-by: Liran Alon <liran.alon@xxxxxxxxxx>
> ---
>  arch/x86/kvm/vmx/vmx.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 31ce6bc2c371..18b0bee662a5 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -6141,7 +6141,10 @@ static int vmx_sync_pir_to_irr(struct kvm_vcpu *vcpu)
>  
>  static bool vmx_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
>  {
> -	return pi_test_on(vcpu_to_pi_desc(vcpu));
> +	struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
> +
> +	return pi_test_on(pi_desc) ||
> +		!bitmap_empty((unsigned long *)pi_desc->pir, NR_VECTORS);
>  }
>  
>  static void vmx_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)

Should we check the bitmap only if SN is false?  We have a precondition
that if SN is clear then non-empty PIR implies ON=1 (modulo the small
window in vmx_vcpu_pi_load of course), so that'd be a bit faster.

Paolo