Re: [PATCH v7 09/20] KVM: selftests: Add VFIO device support to eventfd IRQ test

[Sean Christopherson <seanjc@google.com>]

On Mon, Jun 15, 2026, David Matlack wrote:
> On Mon, Jun 15, 2026 at 11:05 AM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Mon, Jun 15, 2026, David Matlack wrote:
> > > On Fri, Jun 12, 2026 at 5:35 PM <sashiko-bot@kernel.org> wrote:
> > > > > diff --git a/tools/testing/selftests/kvm/irq_test.c b/tools/testing/selftests/kvm/irq_test.c
> > > > > index 9f8895b898214..6888be54ee4a9 100644
> > > > > --- a/tools/testing/selftests/kvm/irq_test.c
> > > > > +++ b/tools/testing/selftests/kvm/irq_test.c
> > > > [ ... ]
> > > > > @@ -55,6 +58,36 @@ static void *vcpu_thread_main(void *arg)
> > > > >       return NULL;
> > > > >  }
> > > > >
> > > > > +static int vfio_setup_msi(struct vfio_pci_device *device)
> > > > > +{
> > > > > +     const int flags = MAP_SHARED | MAP_ANONYMOUS;
> > > > > +     const int prot = PROT_READ | PROT_WRITE;
> > > > > +     struct dma_region *region;
> > > > > +
> > > > > +     /* A driver is required to generate an MSI. */
> > > > > +     TEST_REQUIRE(device->driver.ops);
> > > > > +
> > > > > +     /* Set up a DMA-able region for the driver to use. */
> > > > > +     region = &device->driver.region;
> > > > > +     region->iova = 0;
> > > >
> > > > [Severity: Medium]
> > > > Does hardcoding the DMA region's IOVA to 0 cause IOMMU mapping failures on
> > > > platforms that reserve or restrict IOVA 0?
> > > >
> > > > If IOVA 0 is unsupported, it seems like iommu_map() will fail and trigger a
> > > > test assertion crash instead of running.
> > >
> > > This is probably fine for now given the test is only for x86_64. But
> > > if you are going to send a v8 you can use the VFIO selftests
> > > iova_allocator to fix this, instead of hard-coding.
> >
> > I'm confused, why is this test setting up a DMA-able region in the first place?
> 
> There might be some devices with which you can trigger an MSI with
> pure MMIO. But most devices you need to at least set up some
> descriptors in memory that describe what action you want the device to
> take and then use MMIO to poke the device to read the desriptor and
> perform the action. Reading those descriptors is DMA.
> 
> For Intel DSA specifically, send_msi() works by telling the device to
> perform an 8-byte memcpy and send an interrupt when it's done. The
> driver region contains the descriptor for that memcpy, as well as the
> 8-byte source/destination buffers the device should use.

Ah, gotcha.  For a comment, how's this look?

diff --git a/tools/testing/selftests/kvm/irq_test.c b/tools/testing/selftests/kvm/irq_test.c
index 6888be54ee4a..9a495fdb52f5 100644
--- a/tools/testing/selftests/kvm/irq_test.c
+++ b/tools/testing/selftests/kvm/irq_test.c
@@ -67,7 +67,12 @@ static int vfio_setup_msi(struct vfio_pci_device *device)
        /* A driver is required to generate an MSI. */
        TEST_REQUIRE(device->driver.ops);
 
-       /* Set up a DMA-able region for the driver to use. */
+       /*
+        * Set up a DMA-able region for the driver to use.   Very few devices
+        * provide a way to arbitrarily send interrupts (MSIs), e.g. by writing
+        * an MMIO register.  Instead, most devices send MSIs when an action is
+        * completed, and practically all actions involve DMA of some form.
+        */
        region = &device->driver.region;
        region->iova = 0;
        region->size = SZ_2M;