Re: [PATCH 2/2] KVM: x86: Fix split-irqchip vs interrupt injection window request

[Paolo Bonzini <pbonzini@redhat.com>]

On 14/04/21 04:28, Lai Jiangshan wrote:
> On Tue, Apr 13, 2021 at 8:15 PM Paolo Bonzini <pbonzini@redhat.com> wrote:
>>
>> On 13/04/21 13:03, Lai Jiangshan wrote:
>>> This patch claims that it has a place to
>>> stash the IRQ when EFLAGS.IF=0, but inject_pending_event() seams to ignore
>>> EFLAGS.IF and queues the IRQ to the guest directly in the first branch
>>> of using "kvm_x86_ops.set_irq(vcpu)".
>>
>> This is only true for pure-userspace irqchip.  For split-irqchip, in
>> which case the "place to stash" the interrupt is
>> vcpu->arch.pending_external_vector.
>>
>> For pure-userspace irqchip, KVM_INTERRUPT only cares about being able to
>> stash the interrupt in vcpu->arch.interrupt.injected.  It is indeed
>> wrong for userspace to call KVM_INTERRUPT if the vCPU is not ready for
>> interrupt injection, but KVM_INTERRUPT does not return an error.
> 
> Thanks for the reply.
> 
> May I ask what is the correct/practical way of using KVM_INTERRUPT ABI
> for pure-userspace irqchip.
> 
> gVisor is indeed a pure-userspace irqchip, it will call KVM_INTERRUPT
> when kvm_run->ready_for_interrupt_injection=1 (along with other conditions
> unrelated to our discussion).
> 
> https://github.com/google/gvisor/blob/a9441aea2780da8c93da1c73da860219f98438de/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go#L105
> 
> if kvm_run->ready_for_interrupt_injection=1 when expection pending or
> EFLAGS.IF=0, it would be unexpected for gVisor.

Not with EFLAGS.IF=0.  For pending exception, there is code to handle it 
in inject_pending_event:

         ... if (!vcpu->arch.exception.pending) {
                 if (vcpu->arch.nmi_injected) {
                         static_call(kvm_x86_set_nmi)(vcpu);
                         can_inject = false;
                 } else if (vcpu->arch.interrupt.injected) {
                         static_call(kvm_x86_set_irq)(vcpu);
                         can_inject = false;
                 }
         }
	...
         if (vcpu->arch.exception.pending) {
		...
                 can_inject = false;
         }
	// this is vcpu->arch.interrupt.injected for userspace LAPIC
         if (kvm_cpu_has_injectable_intr(vcpu)) {
                 r = can_inject ? 
static_call(kvm_x86_interrupt_allowed)(vcpu, true) : -EBUSY;
		if (r < 0)
			goto busy;
		...
	}

so what happens is:

- the interrupt will not be injected before the exception

- KVM will schedule an immediate vmexit to inject the interrupt as well

- if (as is likely) the exception has turned off interrupts, the next 
call to inject_pending_event will reach 
static_call(kvm_x86_enable_irq_window) and the interrupt will only be 
injected when IF becomes 1 again.

Paolo