[ewg] Re: Continue of "defer skb_orphan() until irqs enabled"

[Roland Dreier]

 > Another possibility is to simply remove the call to skb_orphan() 
 > from ipoib_send(). Eventually the skb->destructor method will be 
 > called in net_tx_action() (in bottom half context).

That leads to stalls/terrible performance in some cases, because we do
not collect TX completions often enough to free socket space to allow
more packets to be queued so we can get very long stalls.

The best solution would be to get rid of the IRQ-disabling locking (and
stop doing LLTX).  That may be possible without too much pain.