vhost_net: correctly check tx avail during rx busy polling
authorJason Wang <jasowang@redhat.com>
Tue, 5 Sep 2017 01:22:05 +0000 (09:22 +0800)
committerDavid S. Miller <davem@davemloft.net>
Tue, 5 Sep 2017 21:47:32 +0000 (14:47 -0700)
We check tx avail through vhost_enable_notify() in the past which is
wrong since it only checks whether or not guest has filled more
available buffer since last avail idx synchronization which was just
done by vhost_vq_avail_empty() before. What we really want is checking
pending buffers in the avail ring. Fix this by calling
vhost_vq_avail_empty() instead.

This issue could be noticed by doing netperf TCP_RR benchmark as
client from guest (but not host). With this fix, TCP_RR from guest to
localhost restores from 1375.91 trans per sec to 55235.28 trans per
sec on my laptop (Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz).

Fixes: 030881372460 ("vhost_net: basic polling support")
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/vhost/net.c

index 8d2bcae53a2ec9ea1c876625e581bcd429abe365..58585ec8699e8d46100ff5f5e175ab39a7842154 100644 (file)
@@ -635,8 +635,13 @@ static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 
                preempt_enable();
 
-               if (vhost_enable_notify(&net->dev, vq))
+               if (!vhost_vq_avail_empty(&net->dev, vq))
                        vhost_poll_queue(&vq->poll);
+               else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
+                       vhost_disable_notify(&net->dev, vq);
+                       vhost_poll_queue(&vq->poll);
+               }
+
                mutex_unlock(&vq->mutex);
 
                len = peek_head_len(rvq, sk);