From owner-freebsd-hackers  Mon Jun 16 06:50:57 1997
Return-Path: <owner-hackers>
Received: (from root@localhost)
          by hub.freebsd.org (8.8.5/8.8.5) id GAA06044
          for hackers-outgoing; Mon, 16 Jun 1997 06:50:57 -0700 (PDT)
Received: from dg-rtp.dg.com (dg-rtp.rtp.dg.com [128.222.1.2])
          by hub.freebsd.org (8.8.5/8.8.5) with SMTP id GAA06032
          for <hackers@FreeBSD.ORG>; Mon, 16 Jun 1997 06:50:52 -0700 (PDT)
Received: by dg-rtp.dg.com (5.4R3.10/dg-rtp-v02)
	id AA07204; Mon, 16 Jun 1997 09:50:11 -0400
Received: from ponds by dg-rtp.dg.com.rtp.dg.com; Mon, 16 Jun 1997 09:50 EDT
Received: from lakes.water.net (lakes [10.0.0.3]) by ponds.water.net (8.8.5/8.7.3) with ESMTP id OAA23833; Sun, 15 Jun 1997 14:48:01 -0400 (EDT)
Received: (from rivers@localhost) by lakes.water.net (8.8.5/8.6.9) id OAA10921; Sun, 15 Jun 1997 14:56:04 -0400 (EDT)
Date: Sun, 15 Jun 1997 14:56:04 -0400 (EDT)
From: Thomas David Rivers <ponds!rivers@dg-rtp.dg.com>
Message-Id: <199706151856.OAA10921@lakes.water.net>
To: ponds!hq.icb.chel.su!babkin, ponds!time.cdrom.com!jkh
Subject: Re: C optimizer bug ?
Cc: ponds!FreeBSD.ORG!hackers
Content-Type: text
Sender: owner-hackers@FreeBSD.ORG
X-Loop: FreeBSD.org
Precedence: bulk

> > 
> > > I tried everything and finally decided to build the kernel
> > > without the optimization. Everything started working fine!
> > > So the C optimizer seems like having a bug.
> > 
> > Without trying to reduce this down to something others can actually
> > reproduce, however, this basically goes into the same category as
> > UFO sightings and alien abductions.  There's strong evidence that
> > it _might_ happen, but with no concrete proof we can only continue
> > to speculate as to whether or not they're real. :-)
> 
> Okay, here is the description: The function scsi_done() must check
> whether xs->bp is zero and then depending on it call biodone()
> or wakeup(xs). Sometimes it confuses these cases and does the
> wrong thing, in this case the process hangs on tsleep(...,"scsicmd")
> or waiting for biodone() (I can't remember the argument of
> tsleep() for biodone exactly). The probability of hangup on missing
> wakeup(xs) is about 1/4. The hangups on missing biodone() are
> rare.
> 
> In case of any delays caused by printf() or tcpdump or
> using a loopback Ethernet card simulator
> the SCSI simulator is able to finish the operation before the
> scsi_scsi_cmd() function gets control back from the SCSI
> card driver and in this case it just does not sleep. The hangups
> on missing biodone() still occur even with delays.
> 
> I don't know how to reproduce this bug without all those
> simulators but I think it should be possible.


 Hey!!!  

  This sounds an awful lot like my "daily panic" problem - when I
put printf()s around scsi_done, etc... the problem goes away. [Although,
my problem isn't limited to SCSI...]

  Since I've demonstrated this with older and newer versions of FreeBSD,
which used at least two different GCC compilers - I tend to think
it's not a compiler problem...

  I'll bet you'll discover something else is trashing some memory.
Likely overwriting memory that's in a different location when you
don't optimize.  That also why adding printf()s "helps" the problem;
as the printf()s change the size of the stack...

	- Dave Rivers -
> 
> -SB
>