Re: [Qemu-devel] Regression: virtio-pci: convert to ioeventfd callbacks

[Peter Lieven]

Am 28.06.2016 um 09:42 schrieb Cornelia Huck:
> On Tue, 28 Jun 2016 10:03:21 +0300
> "Michael S. Tsirkin" <address@hidden> wrote:
>
> > I notice cleanup is a bit weird:
> >
> >          virtio_queue_set_host_notifier_fd_handler(vq, false, false);
> >          k->ioeventfd_assign(proxy, notifier, n, assign);
> >          event_notifier_cleanup(notifier);
> >
> > I think virtio_queue_set_host_notifier_fd_handler should happen
> > after ioeventfd_assign for symmetry with init?
> Looking at the pre-rework code, ccw used the order now in common code,
> while pci and mmio used the order you suggest.
>
> "Switch the handler back, then unassign the transport's ioeventfd
> backing" made more sense to me (regardless of symmetry) - but we might
> lose a notification?
>
> Peter: Can you check whether your problem goes away if you switch the
> two lines around?

The problem goes away, but its horribly slow. Maybe the lost notifications
you were thinking off.

diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c
index 1313760..7924a59 100644
--- a/hw/virtio/virtio-bus.c
+++ b/hw/virtio/virtio-bus.c
@@ -176,9 +176,9 @@ static int set_host_notifier_internal(DeviceState *proxy, 
VirtioBusState *bus,
             return r;
         }
     } else {
-        virtio_queue_set_host_notifier_fd_handler(vq, false, false);
         k->ioeventfd_assign(proxy, notifier, n, assign);
         event_notifier_cleanup(notifier);
+        virtio_queue_set_host_notifier_fd_handler(vq, false, false);
     }
     return r;
 }

Peter