virtio-net: make rx buf size estimation works for XDP
authorJason Wang <jasowang@redhat.com>
Fri, 23 Dec 2016 14:37:29 +0000 (22:37 +0800)
committerDavid S. Miller <davem@davemloft.net>
Fri, 23 Dec 2016 18:48:55 +0000 (13:48 -0500)
We don't update ewma rx buf size in the case of XDP. This will lead
underestimation of rx buf size which causes host to produce more than
one buffers. This will greatly increase the possibility of XDP page
linearization.

Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/virtio_net.c

index 0778dc8..77ae358 100644 (file)
@@ -584,10 +584,12 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                                put_page(page);
                                head_skb = page_to_skb(vi, rq, xdp_page,
                                                       0, len, PAGE_SIZE);
+                               ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
                                return head_skb;
                        }
                        break;
                case XDP_TX:
+                       ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
                        if (unlikely(xdp_page != page))
                                goto err_xdp;
                        rcu_read_unlock();
@@ -596,6 +598,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                default:
                        if (unlikely(xdp_page != page))
                                __free_pages(xdp_page, 0);
+                       ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
                        goto err_xdp;
                }
        }