virtio_net: avoid (most) NETDEV_TX_BUSY by stopping queue early.
Rusty Russell [Thu, 24 Sep 2009 15:59:20 +0000 (09:59 -0600)]
Now we can tell the theoretical capacity remaining in the output
queue, virtio_net can waste entries by stopping the queue early.

It doesn't work in the case of indirect buffers and kmalloc failure,
but that's rare (we could drop the packet in that case, but other
drivers return TX_BUSY for similar reasons).

For the record, I think this patch reflects poorly on the linux
network API.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Dinesh Subhraveti <dineshs@us.ibm.com>

drivers/net/virtio_net.c

index 420388a..effe8c6 100644 (file)
@@ -1,4 +1,4 @@
-/* A simple network driver using virtio.
+/* A network driver using virtio.
  *
  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
  *
@@ -73,6 +73,7 @@ struct skb_vnet_hdr {
                struct virtio_net_hdr hdr;
                struct virtio_net_hdr_mrg_rxbuf mhdr;
        };
+       unsigned int num_sg;
 };
 
 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
@@ -442,23 +443,24 @@ again:
        return received;
 }
 
-static void free_old_xmit_skbs(struct virtnet_info *vi)
+static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 {
        struct sk_buff *skb;
-       unsigned int len;
+       unsigned int len, tot_sgs = 0;
 
        while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
                pr_debug("Sent skb %p\n", skb);
                __skb_unlink(skb, &vi->send);
                vi->dev->stats.tx_bytes += skb->len;
                vi->dev->stats.tx_packets++;
+               tot_sgs += skb_vnet_hdr(skb)->num_sg;
                kfree_skb(skb);
        }
+       return tot_sgs;
 }
 
 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 {
-       int num;
        struct scatterlist sg[2+MAX_SKB_FRAGS];
        struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
        const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
@@ -502,13 +504,14 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
        else
                sg_set_buf(sg, &hdr->hdr, sizeof(hdr->hdr));
 
-       num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
-       return vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
+       hdr->num_sg = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
+       return vi->svq->vq_ops->add_buf(vi->svq, sg, hdr->num_sg, 0, skb);
 }
 
 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
        struct virtnet_info *vi = netdev_priv(dev);
+       int capacity;
 
 again:
        /* Free up any pending old buffers before queueing new ones. */
@@ -516,27 +519,40 @@ again:
 
        /* Put new one in send queue and do transmit */
        __skb_queue_head(&vi->send, skb);
-       if (likely(xmit_skb(vi, skb) >= 0)) {
-               vi->svq->vq_ops->kick(vi->svq);
-               /* Don't wait up for transmitted skbs to be freed. */
-               skb_orphan(skb);
-               nf_reset(skb);
-               return NETDEV_TX_OK;
+       capacity = xmit_skb(vi, skb);
+
+       /* This can happen with OOM and indirect buffers. */
+       if (unlikely(capacity < 0)) {
+               netif_stop_queue(dev);
+               dev_warn(&dev->dev, "Unexpected full queue\n");
+               if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
+                       vi->svq->vq_ops->disable_cb(vi->svq);
+                       netif_start_queue(dev);
+                       goto again;
+               }
+               return NETDEV_TX_BUSY;
        }
 
-       /* 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
-        * means some were used in the meantime. */
-       if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
-               vi->svq->vq_ops->disable_cb(vi->svq);
-               netif_start_queue(dev);
-               goto again;
+       vi->svq->vq_ops->kick(vi->svq);
+       /* Don't wait up for transmitted skbs to be freed. */
+       skb_orphan(skb);
+       nf_reset(skb);
+
+       /* Apparently nice girls don't return TX_BUSY; stop the queue
+        * before it gets out of hand.  Naturally, this wastes entries. */
+       if (capacity < 2+MAX_SKB_FRAGS) {
+               netif_stop_queue(dev);
+               if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
+                       /* More just got used, free them then recheck. */
+                       capacity += free_old_xmit_skbs(vi);
+                       if (capacity >= 2+MAX_SKB_FRAGS) {
+                               netif_start_queue(dev);
+                               vi->svq->vq_ops->disable_cb(vi->svq);
+                       }
+               }
        }
-       return NETDEV_TX_BUSY;
+
+       return NETDEV_TX_OK;
 }
 
 static int virtnet_set_mac_address(struct net_device *dev, void *p)