RE: [PATCH 2/3] kvm: x86: do not use KVM_REQ_EVENT for APICv interrupt injection
From: Wu, Feng
Date: Wed Sep 28 2016 - 08:07:04 EST
> -----Original Message-----
> From: Paolo Bonzini [mailto:pbonzini@xxxxxxxxxx]
> Sent: Wednesday, September 28, 2016 7:50 PM
> To: Wu, Feng <feng.wu@xxxxxxxxx>; Michael S. Tsirkin <mst@xxxxxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; kvm@xxxxxxxxxxxxxxx;
> yang.zhang.wz@xxxxxxxxx; rkrcmar@xxxxxxxxxx
> Subject: Re: [PATCH 2/3] kvm: x86: do not use KVM_REQ_EVENT for APICv
> interrupt injection
> On 28/09/2016 13:40, Wu, Feng wrote:
> > IIUIC, the issue you describe above is that IPI for posted-interrupts may be
> > issued between
> > vcpu->mode = IN_GUEST_MODE;
> > and
> > local_irq_disable();
> > But if that really happens, we will call kvm_vcpu_kick() in
> > vmx_deliver_posted_interrupt(), hence the vcpu->mode will be changed
> > to EXITING_GUEST_MODE, then we will goto cancel_injection in
> > vcpu_enter_guest, so the posted-interrupt will be delivered to guest
> > in the next vmentry. Seems I cannot see the problem. Do I miss something?
> No, if that happens kvm_trigger_posted_interrupt returns true, hence
> kvm_vcpu_kick is not called.
Oops, I missed the "!" before the function call ...
> With the fix, the IPI is processed as soon
> as the guest enters non-root mode, and the interrupt is injected.
> The other issue occurs when the IPI is sent between
> vcpu->mode = IN_GUEST_MODE;
> In this case, kvm_vcpu_kick is called but it (correctly) doesn't do
> anything because it sees vcpu->mode == OUTSIDE_GUEST_MODE. Then the
> guest is entered with PIR.ON, but the PI interrupt is not pending and
> hence the interrupt is never delivered to the guest.
Why "never", at least, the interrupt should be delivered to the guest in the next
vm-entry, right? I mean vm-entry -> vm-exit -> _vm-entry_ (interrupts will be
delivered at this vm-entery).
> The fix for this
> is to move the RVI update after IN_GUEST_MODE. Then the source CPU uses
> the posted interrupt IPI instead of kvm_cpu_kick, and everything works.