Re: [PATCH 3/9] kvm: qemu: Remove virtio_net tx ring-full heuristic

[Dor Laor <dor.laor@qumranet.com>]

Mark McLoughlin wrote:
> virtio_net tries to guess when it has received a tx
> notification from the guest whether it indicates that the
> guest has no more room in the tx ring and it should
> immediately flush the queued buffers.
>
> The heuristic is based on the fact that there are 128
> buffer entries in the ring and each packet uses 2 buffers
> (i.e. the virtio_net_hdr and the packet's linear data).
>
> Using GSO or increasing the size of the rings will break
> that heuristic, so let's remove it and assume that any
> notification from the guest after we've disabled
> notifications indicates that we should flush our buffers.
>
> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
> ---
>  qemu/hw/virtio-net.c |    3 +--
>  1 files changed, 1 insertions(+), 2 deletions(-)
>
> diff --git a/qemu/hw/virtio-net.c b/qemu/hw/virtio-net.c
> index 31867f1..4adfa42 100644
> --- a/qemu/hw/virtio-net.c
> +++ b/qemu/hw/virtio-net.c
> @@ -175,8 +175,7 @@ static void virtio_net_handle_tx(VirtIODevice *vdev, VirtQueue *vq)
>  {
>      VirtIONet *n = to_virtio_net(vdev);
>  
> -    if (n->tx_timer_active &&
> -	(vq->vring.avail->idx - vq->last_avail_idx) == 64) {
> +    if (n->tx_timer_active) {
>  	vq->vring.used->flags &= ~VRING_USED_F_NO_NOTIFY;
>  	qemu_del_timer(n->tx_timer);
>  	n->tx_timer_active = 0;
>   
Actually we can improve latency a bit more by using this timer only for 
high throughput
scenario. For example, if during the previous timer period no/few 
packets were accumulated,
we can set the flag off and not issue new timer. This way we'll get 
notified immediately without timer
latency. When lots of packets will be transmitted, we'll go back to this 
batch mode again.
Cheers, Dor