David Miller wrote:

 	/* run TX completion thread */
 	if (sblk->idx[0].tx_consumer != tp->tx_cons) {
-		tg3_tx(tp);
+		netif_tx_lock(netdev);
+		__tg3_tx(tp);
+		netif_tx_unlock(netdev);

This will lead to more and longer lock contentions between
->hard_start_xmit() and ->poll() in the normal fast path.  Why
do we need to widen the scope of netif_tx_lock() in this case?

By not taking the netif_tx_lock() during tx completions until
the last moment and only when necessary improves the performance
substantially.