I've also noticed that the number of hard_start_xmit failures 
in e1000 has increased significantly in version 5.1.13-k1. In 
version 5.0.43-k1 the number of failures was much smaller.
    

Interesting.  Felix, would you undo the change[1] below in 5.1.13-k1 and
see what happens?  With the change below, 5.1.13 would be more
aggressive on Tx cleanup, so we'll be quicker waking the queue than
before. 

-scott

        for(i = 0; i < E1000_MAX_INTR; i++)
-               if(!e1000_clean_rx_irq(adapter) &&
+               if(!e1000_clean_rx_irq(adapter) &
                   !e1000_clean_tx_irq(adapter))
                        break;

[1] Something still bothers me about this new form where we're mixing a
bit-wise operator with logical operands.  Should this bother me?