(Note: this list is stale.)

List of easy optimization targets.

- Introduced an unlocked read of the so_comp queue in accept() to
  see if there are items in the queue or not before doing the harder
  work.  If non-blocking, we can return immediately.  This may help
  applications like qmail which appear to do substantial polling of
  the accept queue.

- Explore whether we see real benefits from the current break-out of
  socket locking: we have three primitives (SOCK_LOCK(so),
  SOCKBUF_LOCK(&so->so_rcv), SOCKBUF_LOCK(&so->so_snd)), and they
  could be coallesced or further broken out.

- Investigate relationship between polling at SMP; in particular,
  investigate binding particular polling threads to particular
  interfaces and avoiding undue CPU migration during polling.  In order
  to maintain ordering, it is likely desirable to map each interface
  to only one thread to serialize packet input from the interface.

- Investigate breaking out pcbinfo inpcb chains based on various
  factors, including hashing, in order to reduce contention on pcbinfo
  chain locks.  Also consider reference counting models for inpcbs.  Note
  that the pcbinfo locks are sometimes used to protect so_pcb.