]> git.baikalelectronics.ru Git - kernel.git/commitdiff
virtio-net: lower min ring num_free for efficiency
author? jiang <jiangkidd@hotmail.com>
Tue, 20 Aug 2019 02:51:23 +0000 (02:51 +0000)
committerMichael S. Tsirkin <mst@redhat.com>
Wed, 4 Sep 2019 10:21:17 +0000 (06:21 -0400)
This change lowers ring buffer reclaim threshold from 1/2*queue to budget
for better performance. According to our test with qemu + dpdk, packet
dropping happens when the guest is not able to provide free buffer in
avail ring timely with default 1/2*queue. The value in the patch has been
tested and does show better performance.

Test setup: iperf3 to generate packets to guest (total 30mins, pps 400k, UDP)
avg packets drop before: 2842
avg packets drop after: 360(-87.3%)

Further, current code suffers from a starvation problem: the amount of
work done by try_fill_recv is not bounded by the budget parameter, thus
(with large queues) once in a while userspace gets blocked for a long
time while queue is being refilled. Trigger refills earlier to make sure
the amount of work to do is limited.

Signed-off-by: jiangkidd <jiangkidd@hotmail.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
drivers/net/virtio_net.c

index 4f3de0ac8b0b44a91f06f1c2dd37fde7bc5e0069..ba98e0971b842df0a6e2c690773eb3ee32986b47 100644 (file)
@@ -1331,7 +1331,7 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
                }
        }
 
-       if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
+       if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
                if (!try_fill_recv(vi, rq, GFP_ATOMIC))
                        schedule_delayed_work(&vi->refill, 0);
        }