IPoIB: Recycle loopback skbs instead of freeing and reallocating
authorRoland Dreier <rolandd@cisco.com>
Tue, 10 Jul 2007 20:43:53 +0000 (13:43 -0700)
committerRoland Dreier <rolandd@cisco.com>
Tue, 10 Jul 2007 20:43:53 +0000 (13:43 -0700)
InfiniBand HCAs replicate multicast packets back to the QP that sent
them if that QP is attached to the destination multicast group.  This
means that IPoIB multicasts are often replicated back to the receive
queue of the interface that generated them.  To avoid confusing the
network stack, we drop these duplicates within the IPoIB driver.

However, there's no reason to free the skb that received the duplicate
and then immediately allocate a new skb to post to the receive queue.
We can be more efficient and just repost the same skb.

Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/ulp/ipoib/ipoib_ib.c

index 8404f05..1094488 100644 (file)
@@ -196,6 +196,13 @@ static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
                return;
        }
 
+       /*
+        * Drop packets that this interface sent, ie multicast packets
+        * that the HCA has replicated.
+        */
+       if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
+               goto repost;
+
        /*
         * If we can't allocate a new RX buffer, dump
         * this packet and reuse the old buffer.
@@ -213,24 +220,18 @@ static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
        skb_put(skb, wc->byte_len);
        skb_pull(skb, IB_GRH_BYTES);
 
-       if (wc->slid != priv->local_lid ||
-           wc->src_qp != priv->qp->qp_num) {
-               skb->protocol = ((struct ipoib_header *) skb->data)->proto;
-               skb_reset_mac_header(skb);
-               skb_pull(skb, IPOIB_ENCAP_LEN);
+       skb->protocol = ((struct ipoib_header *) skb->data)->proto;
+       skb_reset_mac_header(skb);
+       skb_pull(skb, IPOIB_ENCAP_LEN);
 
-               dev->last_rx = jiffies;
-               ++priv->stats.rx_packets;
-               priv->stats.rx_bytes += skb->len;
+       dev->last_rx = jiffies;
+       ++priv->stats.rx_packets;
+       priv->stats.rx_bytes += skb->len;
 
-               skb->dev = dev;
-               /* XXX get correct PACKET_ type here */
-               skb->pkt_type = PACKET_HOST;
-               netif_receive_skb(skb);
-       } else {
-               ipoib_dbg_data(priv, "dropping loopback packet\n");
-               dev_kfree_skb_any(skb);
-       }
+       skb->dev = dev;
+       /* XXX get correct PACKET_ type here */
+       skb->pkt_type = PACKET_HOST;
+       netif_receive_skb(skb);
 
 repost:
        if (unlikely(ipoib_ib_post_receive(dev, wr_id)))