From owner-freebsd-hackers  Fri Jan 31 07:34:23 1997
Return-Path: <owner-hackers>
Received: (from root@localhost)
          by freefall.freebsd.org (8.8.5/8.8.5) id HAA27409
          for hackers-outgoing; Fri, 31 Jan 1997 07:34:23 -0800 (PST)
Received: from eldorado.net-tel.co.uk ([193.122.171.253])
          by freefall.freebsd.org (8.8.5/8.8.5) with SMTP id HAA27330
          for <freebsd-hackers@freebsd.org>; Fri, 31 Jan 1997 07:34:03 -0800 (PST)
From: Andrew.Gordon@net-tel.co.uk
Received: (from root@localhost) by eldorado.net-tel.co.uk (8.6.12/8.6.10) id PAA20686; Fri, 31 Jan 1997 15:31:11 GMT
Received: from "/PRMD=NET-TEL/ADMD=GOLD 400/C=GB/" by net-tel.co.uk
        (Route400-RFCGate); Fri, 31 Jan 97 15:27:15 +0000
X400-Received: by mta "eldorado" in "/PRMD=net-tel/ADMD=gold 400/C=gb/"; 
        Relayed; Fri, 31 Jan 97 15:27:15 +0000
X400-Received: by mta "net-tel cambridge" in "/PRMD=net-tel/ADMD=gold 400/C=gb/";
         Relayed; Fri, 31 Jan 97 15:27:13 +0000
X400-Received: by "/PRMD=NET-TEL/ADMD=Gold 400/C=GB/"; Relayed; 
        Fri, 31 Jan 97 15:27:13 +0000
X400-MTS-Identifier: 
        ["/PRMD=NET-TEL/ADMD=Gold 400/C=GB/";hst:8655-970131152713-4AEF]
X400-Content-Type: P2-1984 (2)
X400-Originator: Andrew.Gordon@net-tel.co.uk
Original-Encoded-Information-Types: IA5-Text
X400-Recipients: non-disclosure:;
Date: Fri, 31 Jan 97 15:27:13 +0000
X400-Content-Identifier: Re(2): Re(2): Us
Message-Id: <"5fb2-970131152702-FFD1*/G=Andrew/S=Gordon/O=NET-TEL Computer
 Systems Ltd/PRMD=NET-TEL/ADMD=Gold 400/C=GB/"@MHS>
To: "Ron G. Minnich" <rminnich@Sarnoff.COM>
Cc: freebsd-hackers@freebsd.org
In-Reply-To: <Pine.SUN.3.91.970131093856.16303C-100000@terra>
Subject: Re(2): Re(2): Using rfork() / threads
Sender: owner-hackers@freebsd.org
X-Loop: FreeBSD.org
Precedence: bulk

> well, i have another thing that was (ah well) offered to freebsd called 
> fastlock(). Fastlock was two things: 
> 1) test and set, which ran at memory bandwidth
> 2) a system call to be called only if the lock failed, 
>    which ran at system call speeds. 

Sounds just the ticket for what I need.

I am quite often writing servers for one network protocol or another,
where the the server holds per-user state for a large number of users,
but the users typically spend much of their time asleep at the keyboard
and hence a process-per-user design is wasteful, but in the absence
of async I/O, a one-process design has insufficient parallelism.
The ideal compromise is a small number of processes picking work off a
queue - but this needs the per-user state to be shared between
the processes and a means of sharing out the requests.

You can do some of this with SysV SHM and semaphores, but it gets
inefficient when one of the pieces of per-user state you would
like to share is an open fd.  rfork() and fastlock() seems a much
more efficient combination, though there is a potential problem with
running out of fds in a large system.