Lazy receiver processing

Goal: Guard against receiver livelock and unfair resource allocation
under heavy network load.

Overview:

* Problem: packet processing is done in the context of the receiving
  process.  Packet receive gets the highest priority, followed by
  protocol processing, followed by application processing.  Under
  heavy load, all cycles are consumed with processing incoming
  packets, which are ultimately dropped (livelock).  Also, processing
  is done in the context of the running process (not the receiving
  process), resulting in unfair resource allocations.

* Approach
  - The single IP queue is replaced with a per-socket queue
  - Demultiplexing to the appropriate queue is done as early as
    possible, preferably on the NIC.
  - Protocol processing is done in the context of the receiving
    application. 

* TCP is more complicated than UDP because all processing cannot occur
  in the receive system call.

Comments:

* The unit of resource control is the socket.  A server could obtain
  more resources by opering more sockets.

* It isn't clear how this would respond to spoofed connections.  In
  the worst case, setting up per-socket queues could open up a
  denial-of-service attack.

* LRP does not address the issue of sending bandwidth; a single
  sending queue is shared by all sockets.