vnc: 0.841
other: 0.828
semantic: 0.822
device: 0.743
PID: 0.734
debug: 0.729
permissions: 0.702
graphic: 0.701
KVM: 0.695
performance: 0.687
network: 0.597
boot: 0.593
files: 0.576
socket: 0.547

dataplane interrupt handler doesn't support msi

hw/block/dataplane/virtio-blk.c commit 1010cadf62332017648abee0d7a3dc7f2eef9632

in the function notify_guest_bh, the function virtio_notify_irqfd is called
to deliver the interrupt corresponding to the vq

however, without the dataplane, hw/block/virtio_blk_req_complete calls
virtio_notify to deliver the interrupt (immediately). this goes though
a slightly more involved path that calls virtio_pci_notify which includes
a case to handle msi interrupts.

so, msi interrupts with block devices aren't serviced when using dataplane
batching.

diff --git a/hw/block/dataplane/virtio-blk.c b/hw/block/dataplane/virtio-blk.c
index 101f32c..31d9eb8 100644
--- a/hw/block/dataplane/virtio-blk.c
+++ b/hw/block/dataplane/virtio-blk.c
@@ -73,7 +73,7 @@ static void notify_guest_bh(void *opaque)
             unsigned i = j + ctzl(bits);
             VirtQueue *vq = virtio_get_queue(s->vdev, i);

-            virtio_notify_irqfd(s->vdev, vq);
+            virtio_notify(s->vdev, vq);

             bits &= bits - 1; /* clear right-most bit */
         }


oh right, another note. this only manifests when using kvm.


On Mon, May 14, 2018 at 03:00:44AM -0000, eric hoffman wrote:
> diff --git a/hw/block/dataplane/virtio-blk.c b/hw/block/dataplane/virtio-blk.c
> index 101f32c..31d9eb8 100644
> --- a/hw/block/dataplane/virtio-blk.c
> +++ b/hw/block/dataplane/virtio-blk.c
> @@ -73,7 +73,7 @@ static void notify_guest_bh(void *opaque)
>              unsigned i = j + ctzl(bits);
>              VirtQueue *vq = virtio_get_queue(s->vdev, i);
> 
> -            virtio_notify_irqfd(s->vdev, vq);
> +            virtio_notify(s->vdev, vq);
> 
>              bits &= bits - 1; /* clear right-most bit */
>          }

Please send patches to <email address hidden>.  Guidelines for submitting
patches are here:
https://wiki.qemu.org/Contribute/SubmitAPatch

The issue with this approach is that hw/pci/msi.c:msi_send_message()
invokes device emulation outside the QEMU global mutex (it calls into
the APIC to send MSIs).  I've CCed Paolo Bonzini to check whether doing
this is thread-safe.

Stefan


thanks for looking at this Stefan - since I don't have any context of exactly the kind of environmental issues like threading, the patch posted here isn't really a suggested fix.

it does in general seem helpful if batched interrupts have the same delivery semantics as non-deferred. 

This bug is invalid. MSI/MSI-X interrupts are properly serviced when dataplane batching is used. The original problem was in the incorrect virtio driver initialization sequence (virtqueues and MSI-X interrupts configured after DRIVER_OK bit is set).

This bug can be closed as INVALID.