On Thu, 18 Dec 2014 19:25:19 -0800
Raghu Vatsavayi [off-list ref] wrote:

+static int liquidio_xmit(struct sk_buff *skb, struct net_device *netdev)
+{
+	struct lio *lio;
+	struct octnet_buf_free_info *finfo;
+	union octnic_cmd_setup cmdsetup;
+	struct octnic_data_pkt ndata;
+	struct octeon_device *oct;
+	int cpu = 0, status = 0;
+	int q_idx = 0;
+	int xmit_more;
+
+	lio = GET_LIO(netdev);
+	oct = lio->oct_dev;
+	atomic64_inc((atomic64_t *)&lio->stats.tx_packets);
+	lio->stats.tx_bytes += skb->len;
+
+	if (!ifstate_check(lio, LIO_IFSTATE_TXENABLED)) {
+		lio_info(lio, tx_err, "Transmit Busy, not enabled\n");
+		return NETDEV_TX_BUSY;
+	}

This kind of busy handling in transmit, will cause the send in higher
level to retry and spin the CPU. Better to do proper flow control
like other drivers.