* Fix race in Tx performance path with tx_lock.  Between checking
  if we're out of resources and stopping the queue, we can get
  a hard interrupt which will clean up all Tx work, and wake
  the queue.  Coming out of hard interrupt context, we stop the
  queue even though no work was queued, and all work completed
  has been cleaned up.  Scenario requires ring to be completely
  filled, which is more likely to happen with TSO, since each
  TSO send consumes multiple ring entries.