virtio_net: return NETDEV_TX_BUSY instead of queueing an extra skb.
authorRusty Russell <rusty@rustcorp.com.au>
Thu, 24 Sep 2009 15:59:18 +0000 (09:59 -0600)
committerRusty Russell <rusty@rustcorp.com.au>
Thu, 24 Sep 2009 00:29:19 +0000 (09:59 +0930)
This effectively reverts 99ffc696d10b28580fe93441d627cf290ac4484c
"virtio: wean net driver off NETDEV_TX_BUSY".

The complexity of queuing an skb (setting a tasklet to re-xmit) is
questionable, especially once we get rid of the other reason for the
tasklet in the next patch.

If the skb won't fit in the tx queue, just return NETDEV_TX_BUSY.
This is frowned upon, so a followup patch uses a more complex solution.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
drivers/net/virtio_net.c

index dc4c68718976c91b7604aa987990d4757ee5d4ff..222f3d098ae47d04a3f8666f5a16162b3ebb0d41 100644 (file)
@@ -48,9 +48,6 @@ struct virtnet_info
        struct napi_struct napi;
        unsigned int status;
 
-       /* The skb we couldn't send because buffers were full. */
-       struct sk_buff *last_xmit_skb;
-
        /* If we need to free in a timer, this is it. */
        struct timer_list xmit_free_timer;
 
@@ -120,9 +117,8 @@ static void skb_xmit_done(struct virtqueue *svq)
        /* We were probably waiting for more output buffers. */
        netif_wake_queue(vi->dev);
 
-       /* Make sure we re-xmit last_xmit_skb: if there are no more packets
-        * queued, start_xmit won't be called. */
-       tasklet_schedule(&vi->tasklet);
+       if (vi->free_in_tasklet)
+               tasklet_schedule(&vi->tasklet);
 }
 
 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
@@ -543,12 +539,7 @@ static void xmit_tasklet(unsigned long data)
        struct virtnet_info *vi = (void *)data;
 
        netif_tx_lock_bh(vi->dev);
-       if (vi->last_xmit_skb && xmit_skb(vi, vi->last_xmit_skb) >= 0) {
-               vi->svq->vq_ops->kick(vi->svq);
-               vi->last_xmit_skb = NULL;
-       }
-       if (vi->free_in_tasklet)
-               free_old_xmit_skbs(vi);
+       free_old_xmit_skbs(vi);
        netif_tx_unlock_bh(vi->dev);
 }
 
@@ -560,28 +551,16 @@ again:
        /* Free up any pending old buffers before queueing new ones. */
        free_old_xmit_skbs(vi);
 
-       /* If we has a buffer left over from last time, send it now. */
-       if (unlikely(vi->last_xmit_skb) &&
-           xmit_skb(vi, vi->last_xmit_skb) < 0)
-               goto stop_queue;
-
-       vi->last_xmit_skb = NULL;
-
        /* Put new one in send queue and do transmit */
-       if (likely(skb)) {
-               __skb_queue_head(&vi->send, skb);
-               if (xmit_skb(vi, skb) < 0) {
-                       vi->last_xmit_skb = skb;
-                       skb = NULL;
-                       goto stop_queue;
-               }
+       __skb_queue_head(&vi->send, skb);
+       if (likely(xmit_skb(vi, skb) >= 0)) {
+               vi->svq->vq_ops->kick(vi->svq);
+               return NETDEV_TX_OK;
        }
-done:
-       vi->svq->vq_ops->kick(vi->svq);
-       return NETDEV_TX_OK;
 
-stop_queue:
+       /* Ring too full for this packet, remove it from queue again. */
        pr_debug("%s: virtio not prepared to send\n", dev->name);
+       __skb_unlink(skb, &vi->send);
        netif_stop_queue(dev);
 
        /* Activate callback for using skbs: if this returns false it
@@ -591,12 +570,7 @@ stop_queue:
                netif_start_queue(dev);
                goto again;
        }
-       if (skb) {
-               /* Drop this skb: we only queue one. */
-               vi->dev->stats.tx_dropped++;
-               kfree_skb(skb);
-       }
-       goto done;
+       return NETDEV_TX_BUSY;
 }
 
 static int virtnet_set_mac_address(struct net_device *dev, void *p)