On Tue, Jun 17, 2014 at 1:20 PM, David Miller [off-list ref] wrote:

From: Iyappan Subramanian <redacted>
Date: Tue, 17 Jun 2014 12:21:08 -0700

quoted

On Mon, Jun 16, 2014 at 9:42 PM, David Miller [off-list ref] wrote:

quoted

From: Iyappan Subramanian <redacted>
Date: Mon, 16 Jun 2014 17:18:46 -0700

quoted

+static netdev_tx_t xgene_enet_start_xmit(struct sk_buff *skb,
+                                      struct net_device *ndev)
+{
+     struct xgene_enet_pdata *pdata = netdev_priv(ndev);
+     struct xgene_enet_desc_ring *tx_ring = pdata->tx_ring;
+     struct xgene_enet_desc_ring *cp_ring = tx_ring->cp_ring;
+     u32 tx_level, cq_level;
+
+     tx_level = xgene_enet_ring_len(tx_ring);
+     cq_level = xgene_enet_ring_len(cp_ring);
+     if (unlikely(tx_level > pdata->tx_qcnt_hi ||
+                  cq_level > pdata->cp_qcnt_hi)) {
+             netif_stop_queue(ndev);
+             return NETDEV_TX_BUSY;
+     }
+
+     if (xgene_enet_setup_tx_desc(tx_ring, skb))
+             return NETDEV_TX_OK;

If you return NETDEV_TX_OK, it is your responsibility to deal with the
SKB, because you own it.

In particular, you have to free the packet.

skb freeing happens in xgene_enet_tx_completion().

You're not queueing up the packet when DMA mapping the packet fails,
which is the only condition that causes xgene_enet_setup_tx_desc() to
return an error, so xgene_enet_tx_completion() will not be called.

Can you really not see this?

I am sorry I missed that.  I will fix this.