>>>>> "Warner" == Warner Losh <imp@village.org> writes:
    Warner> How is that different from taking out fine grain locks?

  Well, there are several ways to parallelize network code:
thread-per-packet, thread-per-layer, thread-per-connection,
thread-per-function, and other more "vector" techniques.
  Thread-per-connection means that you just need a coarse-grained lock
on the PCB. Thread-per-packet may spend all its time competing for
locks. Thread-per-layer has similar problems. Stock SVR4 STREAMS
TCP/IP uses the last approach...

  My general bibliography is at http://www.milkyway.com/People/Michael_Richardson/at.bib
  See Nahum*

   :!mcr!:            |  Cow#1: Are you worried about getting Mad Cow Disease?
   Michael Richardson |  Cow#2: No. I'm a duck.
 Home: mcr@sandelman.ocunix.on.ca. PGP key available.