[openib-general] [PATCH/RFC] IB/ipoib: add selective tx signaling

[Roland Dreier]

 > netif_queue_stopped is an atomic, I think that would be sufficient, no?

netif_queue_stopped does test_bit(), which is an implicit memory
barrier.  But I think it's pretty subtle -- you have to look at every
place tx_head and tx_tail are updated and decide if there needs to be
a corresponding barrier.

For example, in:

+	if (netif_queue_stopped(dev) &&
+	    priv->tx_head - priv->tx_tail < ipoib_sendq_size) {

there is a barrier in netif_queue_stopped() but no control in what
order tx_head and tx_tail are read...

 - R.