Re: [PATCH] KVM: nVMX: Rework event injection and recovery

[Gleb Natapov <gleb@redhat.com>]

On Thu, Feb 21, 2013 at 11:33:30AM +0100, Jan Kiszka wrote:
> On 2013-02-21 11:28, Jan Kiszka wrote:
> > On 2013-02-21 11:18, Jan Kiszka wrote:
> >> On 2013-02-21 11:06, Gleb Natapov wrote:
> >>> On Thu, Feb 21, 2013 at 10:43:57AM +0100, Jan Kiszka wrote:
> >>>> On 2013-02-21 10:22, Gleb Natapov wrote:
> >>>>> On Wed, Feb 20, 2013 at 06:50:50PM +0100, Jan Kiszka wrote:
> >>>>>> On 2013-02-20 18:24, Jan Kiszka wrote:
> >>>>>>> On 2013-02-20 18:01, Gleb Natapov wrote:
> >>>>>>>> On Wed, Feb 20, 2013 at 03:37:51PM +0100, Jan Kiszka wrote:
> >>>>>>>>> On 2013-02-20 15:14, Nadav Har'El wrote:
> >>>>>>>>>> Hi,
> >>>>>>>>>>
> >>>>>>>>>> By the way, if you haven't seen my description of why the current code
> >>>>>>>>>> did what it did, take a look at
> >>>>>>>>>> http://www.mail-archive.com/kvm@vger.kernel.org/msg54478.html
> >>>>>>>>>> Another description might also come in handy:
> >>>>>>>>>> http://www.mail-archive.com/kvm@vger.kernel.org/msg54476.html
> >>>>>>>>>>
> >>>>>>>>>> On Wed, Feb 20, 2013, Jan Kiszka wrote about "[PATCH] KVM: nVMX: Rework event injection and recovery":
> >>>>>>>>>>> This aligns VMX more with SVM regarding event injection and recovery for
> >>>>>>>>>>> nested guests. The changes allow to inject interrupts directly from L0
> >>>>>>>>>>> to L2.
> >>>>>>>>>>>
> >>>>>>>>>>> One difference to SVM is that we always transfer the pending event
> >>>>>>>>>>> injection into the architectural state of the VCPU and then drop it from
> >>>>>>>>>>> there if it turns out that we left L2 to enter L1.
> >>>>>>>>>>
> >>>>>>>>>> Last time I checked, if I'm remembering correctly, the nested SVM code did
> >>>>>>>>>> something a bit different: After the exit from L2 to L1 and unnecessarily
> >>>>>>>>>> queuing the pending interrupt for injection, it skipped one entry into L1,
> >>>>>>>>>> and as usual after the entry the interrupt queue is cleared so next time
> >>>>>>>>>> around, when L1 one is really entered, the wrong injection is not attempted.
> >>>>>>>>>>
> >>>>>>>>>>> VMX and SVM are now identical in how they recover event injections from
> >>>>>>>>>>> unperformed vmlaunch/vmresume: We detect that VM_ENTRY_INTR_INFO_FIELD
> >>>>>>>>>>> still contains a valid event and, if yes, transfer the content into L1's
> >>>>>>>>>>> idt_vectoring_info_field.
> >>>>>>>>>>
> >>>>>>>>>>> To avoid that we incorrectly leak an event into the architectural VCPU
> >>>>>>>>>>> state that L1 wants to inject, we skip cancellation on nested run.
> >>>>>>>>>>
> >>>>>>>>>> I didn't understand this last point.
> >>>>>>>>>
> >>>>>>>>> - prepare_vmcs02 sets event to be injected into L2
> >>>>>>>>> - while trying to enter L2, a cancel condition is met
> >>>>>>>>> - we call vmx_cancel_interrupts but should now avoid filling L1's event
> >>>>>>>>>   into the arch event queues - it's kept in vmcs12
> >>>>>>>>>
> >>>>>>>> But what if we put it in arch event queue? It will be reinjected during
> >>>>>>>> next entry attempt, so nothing bad happens and we have one less if() to explain,
> >>>>>>>> or do I miss something terrible that will happen?
> >>>>>>>
> >>>>>>> I started without that if but ran into troubles with KVM-on-KVM (L1
> >>>>>>> locks up). Let me dig out the instrumentation and check the event flow
> >>>>>>> again.
> >>>>>>
> >>>>>> OK, got it again: If we transfer an IRQ that L1 wants to send to L2 into
> >>>>>> the architectural VCPU state, we will also trigger enable_irq_window.
> >>>>>> And that raises KVM_REQ_IMMEDIATE_EXIT again as it thinks L0 wants
> >>>>>> inject. That will send us into an endless loop.
> >>>>>>
> >>>>> Why would we trigger enable_irq_window()? enable_irq_window() triggers
> >>>>> only if interrupt is pending in one of irq chips, not in architectural
> >>>>> VCPU state.
> >>>>
> >>>> Precisely this is the case if an IRQ for L1 arrived while we tried to
> >>>> enter L2 and caused the cancellation above.
> >>>>
> >>> But during next entry the cancelled interrupt is transfered
> >>> from architectural VCPU state to VM_ENTRY_INTR_INFO_FIELD by
> >>> inject_pending_event()->vmx_inject_irq(), so at the point where
> >>> enable_irq_window() is called the state is exactly the same no matter
> >>> whether we canceled interrupt or not during previous entry attempt. What
> >>> am I missing?
> >>
> >> Maybe that we normally either have an external IRQ pending in some IRQ
> >> chip or in the VCPU architectural state, not both at the same time? By
> >> transferring something that doesn't come from a virtual IRQ chip of L0
> >> (but from the one in L1) into the architectural state, we break this
> >> assumption.
> >>
> >>> Oh may be I am missing that if we do not cancel interrupt
> >>> then inject_pending_event() will skip
> >>>   if (vcpu->arch.interrupt.pending)
> >>>     ....
> >>
> >> If we do not cancel, we will not inject at all (due to missing
> >> KVM_REQ_EVENT).
> >>
> >>> and will inject interrupt from APIC that caused cancellation of previous
> >>> entry, but then this is a bug since this new interrupt will overwrite
> >>> the one that is still in VM_ENTRY_INTR_INFO_FIELD from previous entry
> >>> attempt and there may be another pending interrupt in APIC anyway that
> >>> will cause enable_irq_window() too.
> >>
> >> Maybe the issue is that we do not properly simulate a VMEXIT on an
> >> external interrupt during vmrun (like SVM does). Need to check for this
> >> case again...
> > 
> > static int vmx_interrupt_allowed(struct kvm_vcpu *vcpu)
> > {
> > 	if (is_guest_mode(vcpu) && nested_exit_on_intr(vcpu)) {
> > 		struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
> > 		if (to_vmx(vcpu)->nested.nested_run_pending ||
> > 		    (vmcs12->idt_vectoring_info_field &
> > 		     VECTORING_INFO_VALID_MASK))
> > 			return 0;
> > 		nested_vmx_vmexit(vcpu);
> > 		vmcs12->vm_exit_reason = EXIT_REASON_EXTERNAL_INTERRUPT;
> > 		vmcs12->vm_exit_intr_info = 0;
> > 		...
> > 
> > I do not understand ATM why we refuse to simulate a vmexit due to an
> > external interrupt when we are about to run L2 or have something in
> > idt_vectoring_info_field. The external interrupt would not overwrite
> > idt_vectoring_info_field but should end up in vm_exit_intr_info.
> 
> Explained in 51cfe38ea5: idt_vectoring_info_field and vm_exit_intr_info
> must not be valid at the same time.
> 
Interestingly, if we transfer interrupt from idt_vectoring_info into
arch VCPU state we can drop this check because vmx_interrupt_allowed()
will not be called while there is an event to reinject. 51cfe38ea5 still
does not explain why nested_run_pending is needed. We cannot #vmexit
without entering L2, but we can undo VMLAUNCH/VMRESUME emulation leaving
rip pointing to the instruction. We can start by moving
skip_emulated_instruction() from nested_vmx_run() to nested_vmx_vmexit().

--
			Gleb.