[Qemu-devel] [PATCH] virtio-net: enable ioeventfd even if vhost=off

[Paolo Bonzini <pbonzini@redhat.com>]

virtio-net-pci does not enable ioeventfd for historical reasons (and
nobody ever checked whether it should be revisited).  Note that other
backends do enable ioeventfd for virtio-net.

However, it has a major effect on performance.  On Windows, throughput is
_multiplied_ by 2 or 3 on TCP_STREAM (on small packets it is "only" a 30%
improvement) and a little less so on TCP_MAERTS albeit still very much
statistically significant.  Latency also has a single digit improvement.

This is not visible when using vhost, which forces ioeventfd=on, but it
is substantial without vhost.  In addition, also on Windows and with the
RHEL 7.3 kernel, APICv seems to slow down virtio-net performance a bit,
but the penalty with this patch goes from -25% to -7%.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 hw/virtio/virtio-pci.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
index 023a020..0cd42ae 100644
--- a/hw/virtio/virtio-pci.c
+++ b/hw/virtio/virtio-pci.c
@@ -2278,7 +2278,7 @@ static const TypeInfo virtio_serial_pci_info = {
 
 static Property virtio_net_properties[] = {
     DEFINE_PROP_BIT("ioeventfd", VirtIOPCIProxy, flags,
-                    VIRTIO_PCI_FLAG_USE_IOEVENTFD_BIT, false),
+                    VIRTIO_PCI_FLAG_USE_IOEVENTFD_BIT, true),
     DEFINE_PROP_UINT32("vectors", VirtIOPCIProxy, nvectors, 3),
     DEFINE_PROP_END_OF_LIST(),
 };
-- 
2.9.3