virtio: don't enable notifications during polling

Virtqueue notifications are not necessary during polling, so we disable
them.  This allows the guest driver to avoid MMIO vmexits.
Unfortunately the virtio-blk and virtio-scsi handler functions re-enable
notifications, defeating this optimization.

Fix virtio-blk and virtio-scsi emulation so they leave notifications
disabled.  The key thing to remember for correctness is that polling
always checks one last time after ending its loop, therefore it's safe
to lose the race when re-enabling notifications at the end of polling.

There is a measurable performance improvement of 5-10% with the null-co
block driver.  Real-life storage configurations will see a smaller
improvement because the MMIO vmexit overhead contributes less to
latency.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20191209210957.65087-1-stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

1 files changed, 6 insertions, 6 deletions

diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
index 7bc6a9455e..95d8ff8508 100644
--- a/hw/virtio/virtio.c
+++ b/hw/virtio/virtio.c
@@ -432,6 +432,11 @@ static void virtio_queue_packed_set_notification(VirtQueue *vq, int enable)
     }
 }
 
+bool virtio_queue_get_notification(VirtQueue *vq)
+{
+    return vq->notification;
+}
+
 void virtio_queue_set_notification(VirtQueue *vq, int enable)
 {
     vq->notification = enable;
@@ -3410,17 +3415,12 @@ static bool virtio_queue_host_notifier_aio_poll(void *opaque)
 {
     EventNotifier *n = opaque;
     VirtQueue *vq = container_of(n, VirtQueue, host_notifier);
-    bool progress;
 
     if (!vq->vring.desc || virtio_queue_empty(vq)) {
         return false;
     }
 
-    progress = virtio_queue_notify_aio_vq(vq);
-
-    /* In case the handler function re-enabled notifications */
-    virtio_queue_set_notification(vq, 0);
-    return progress;
+    return virtio_queue_notify_aio_vq(vq);
 }
 
 static void virtio_queue_host_notifier_aio_poll_end(EventNotifier *n)