From what I saw I am guessing the root cause was somewhere in the network stack. There were a few network fixes listed in the 5.0 changes webpage including:
-Fixed READ/WRITE confusion in bnx(4)
-Prevent the network stack from enqueueing mbuf(9)s on sockets marked as SS_CANTRCVMORE, as was done for routing sockets.
The second in particular sounds likes my problem. I will probably end up trying both and seeing which works best in my environment.
|