Re: [PATCH v5 07/21] KVM: selftests: Verify IRQ bypass works in IRQ test

[Sean Christopherson <seanjc@google.com>]

On Thu, Jun 04, 2026, Sean Christopherson wrote:
> On Thu, Jun 04, 2026, Sean Christopherson wrote:
> > The shortlog is at best misleading.  There are zero guarantees that IRQ bypass
> > is supported and enabled.
> > 
> > On Thu, Jun 04, 2026, Josh Hilke wrote:
> > > From: David Matlack <dmatlack@google.com>
> > > 
> > > Trigger interrupts from a VFIO device instead of emulating interrupts
> > 
> > It's not emulating, it's synthesizing.  Emulating implies there's a device of
> > some kind that the test is mimicking.
> > 
> > > using KVM eventfds. This verifies that guests receive interrupts via IRQ
> > > bypass.
> > 
> > No, it verifies that delivery of a VFIO MSI-X through VFIO=>KVM works, and *may*
> > verify IRQ bypass.  And all of that very much relies on the eventfds to be in
> > place.
> > 
> > > @@ -119,7 +160,17 @@ int main(int argc, char **argv)
> > >  	vm = vm_create_with_vcpus(nr_vcpus, guest_code, vcpus);
> > >  	vm_install_exception_handler(vm, vector, guest_irq_handler);
> > >  
> > > -	eventfd = kvm_new_eventfd();
> > > +	if (device_bdf) {
> > > +		iommu = iommu_init(default_iommu_mode);
> > 
> > This needs:
> > 
> > diff --git tools/testing/selftests/kvm/irq_test.c tools/testing/selftests/kvm/irq_test.c
> > index cf4568718cee..2e7e100d4815 100644
> > --- tools/testing/selftests/kvm/irq_test.c
> > +++ tools/testing/selftests/kvm/irq_test.c
> > @@ -235,6 +235,8 @@ int main(int argc, char **argv)
> >         }
> >  
> >         if (device_bdf) {
> > +               __open_path_or_exit("/dev/iommu", O_RDONLY, "Is IOMMUFD available?");
> > +
> >                 iommu = iommu_init(default_iommu_mode);
> 
> *sigh* 
> 
> <rant>
> This is beyond frustating.  I have A PERFECTLY FUNCTIONAL KERNEL, but this test
> VERY SUBTLY "defaults" to IOMMUFD.  ARGH!!!!!

Ok, here's what I ended up with.  A param to let the user force a specific type:

	printf("-t	Override the IOMMU type to use (vfio_type1_iommu or iommufd)\n");


And then auto-probe logic to make a halfway decent guess as to which type/mode
to use.

  static const char *probe_iommu_type(void)
  {
	int io_fd;

	io_fd = open("/dev/iommu", O_RDONLY);
	if (io_fd >= 0) {
		close(io_fd);
		return MODE_IOMMUFD;
	}

	io_fd = __open_path_or_exit("/dev/vfio", O_RDONLY,
				    "Is VFIO (or IOMMUFD) loaded and enabled?");
	close(io_fd);
	return MODE_VFIO_TYPE1_IOMMU;
  }

...

		if (!iommu_type)
			iommu_type = probe_iommu_type();

I'll probably throw a TODO in there, because the VFIO/IOMMU infrastructure really
needs to handle this, *especially* since "default_iommu_mode" is globally visible.
Hardcoding a "default" to an entirely optional thing is just evil.

E.g. instead of iommu_init() => lookup_iommu_mode() using default_iommu_mode when
passed a NULL pointer, have it choose the type that's most likely to work.  Then
the IRQ test can drop its probing logic.