> > >> This is not what I asked. You should instead do the check after pi_clear_sn. > > >> > > > > > > I think the SN has been cleared here before test the bitmap. > > > The SN will be set when the vCPU is schedule out. ID: > > > 28b835d60fcc2498e717cf5e6f0c3691c24546f7 > > > But SN will be cleared when sched in. > > > > > > Another place is when vCPU run out of the vcpu_run() function: > > > kvm_arch_vcpu_ioctl_run() > > > vcpu_load(vcpu); -> kvm_arch_vcpu_load -> vmx_vcpu_load -> vmx_vcpu_pi_load -> new.sn = 0; > > > vcpu_run(vcpu); > > > for(;;) > > > vcpu_put(vcpu); -> kvm_arch_vcpu_put -> vmx_vcpu_put -> > > > vmx_vcpu_pi_put -> pi_set_sn() But SN will be cleared in vcpu_load() > > > before back to vcpu_run() > > > > Yes, but you're changing the wrong path. The patch is affecting _all_ vmentries, not just those after PID.SN has been cleared. > > > > As I mentioned in the previous email, KVM relies on the SDM's > > invariant that ON where PID.ON=1 whenever PID.PIR!=0. Invariants are your best friend when dealing with complicated multi-processor > code so I don't want to change that. > > > > It's the VT-d pi_clear_sn path that I want to be changed, because it's > > VT-d and specifically SN that complicates the very simple definition in the SDM. By modifying the pi_clear_sn path, you ensure the > invariant is respected and everyone is happy. > > Hi Paolo, > How about like this: > > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index 820a03b..dfc5e3d 100644 > --- a/arch/x86/kvm/vmx/vmx.c > +++ b/arch/x86/kvm/vmx/vmx.c > @@ -1219,6 +1219,9 @@ static void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu) > new.ndst = (dest << 8) & 0xFF00; > > new.sn = 0; > + > + if (!bitmap_empty((unsigned long *)new.pir, NR_VECTORS)) > + new.on = 1; Sorry, should be: + if (!bitmap_empty((unsigned long *)pi_desc->pir, NR_VECTORS)) Luwei Kang > } while (cmpxchg64(&pi_desc->control, old.control, > new.control) != old.control); } > > Thanks, > Luwei Kang