On Thu, 2014-03-27 at 09:01 -0700, Eric Dumazet wrote:

quoted

+	HARD_TX_LOCK(dev, txq, smp_processor_id());
+	if (!netif_xmit_frozen_or_stopped(txq)) {
+		ret = ops->ndo_start_xmit(skb, dev);
+		if (ret == NETDEV_TX_OK)
+			txq_trans_update(txq);

I think this is problematic.

If you have concurrent traffic going through txq, and a flood
going through packet_direct_xmit(), the timeout will trigger on txq
because we do no longer update txq->trans_start

So I think you should force the update

	txq->trans_start = jiffies;

Oh well, scratch that, I missed that you did the update.