Re: [PATCH v2 02/11] KVM: SVM: Don't BUG if userspace injects a soft interrupt with GIF=0

[Maciej S. Szmigiero]

On 28.04.2022 17:04, Sean Christopherson wrote:
> On Thu, Apr 28, 2022, Maxim Levitsky wrote:
> > On Thu, 2022-04-28 at 15:27 +0200, Maciej S. Szmigiero wrote:
> > > On 28.04.2022 09:35, Maxim Levitsky wrote:
> > > > On Sat, 2022-04-23 at 02:14 +0000, Sean Christopherson wrote:
> > > > > From: Maciej S. Szmigiero <maciej.szmigiero@xxxxxxxxxx>
> > > > >
> > > > > Don't BUG/WARN on interrupt injection due to GIF being cleared if the
> > > > > injected event is a soft interrupt, which are not actually IRQs and thus
> > > >
> > > > Are any injected events subject to GIF set? I think that EVENTINJ just injects
> > > > unconditionaly whatever hypervisor puts in it.
> > >
> > > That's right, EVENTINJ will pretty much always inject, even when the CPU
> > > is in a 'wrong' state (like for example, injecting a hardware interrupt
> > > or a NMI with GIF masked).
> > >
> > > But KVM as a L0 is not supposed to inject a hardware interrupt into guest
> > > with GIF unset since the guest is obviously not expecting it then.
> > > Hence this WARN_ON().
> >
> > If you mean L0->L1 injection, that sure, but if L1 injects interrupt to L2,
> > then it should always be allowed to do so.
>
> Yes, L1 can inject whatever it wants, whenever it wants.
>
> I kept the WARN_ON() under the assumption that KVM would refuse to inject IRQs
> stuffed by userspace if GIF is disabled, but looking at the code again, I have
> no idea why I thought that.  KVM_SET_VCPU_EVENTS blindly takes whatever userspace
> provides, I don't see anything that would prevent userspace from shoving in a
> hardware IRQ.

You both are right, while KVM itself would not inject IRQ with GIF masked and
a nested VMRUN would enable GIF unconditionally, userspace via KVM_SET_VCPU_EVENTS
does not have this restriction.

Will remove this WARN_ON() then.

Thanks,
Maciej