From owner-freebsd-hackers  Sun Mar 14 22: 3:43 1999
Delivered-To: freebsd-hackers@freebsd.org
Received: from cs.rice.edu (cs.rice.edu [128.42.1.30])
	by hub.freebsd.org (Postfix) with ESMTP id 4336114EA8
	for <freebsd-hackers@freebsd.org>; Sun, 14 Mar 1999 22:03:40 -0800 (PST)
	(envelope-from aron@cs.rice.edu)
Received: (from aron@localhost)
	by cs.rice.edu (8.9.0/8.9.0) id AAA01202
	for freebsd-hackers@freebsd.org; Mon, 15 Mar 1999 00:03:21 -0600 (CST)
From: Mohit Aron <aron@cs.rice.edu>
Message-Id: <199903150603.AAA01202@cs.rice.edu>
Subject: bug in performance monitoring code
To: freebsd-hackers@freebsd.org
Date: Mon, 15 Mar 1999 00:03:21 -0600 (CST)
X-Mailer: ELM [version 2.4 PL25]
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: owner-freebsd-hackers@FreeBSD.ORG
Precedence: bulk
X-Loop: FreeBSD.ORG

Hi,
	there's a bug in the performance monitoring code in the file
/sys/i386/i386/perfmon.c that implements ioctl calls to the device
/dev/perfmon. I'm using FreeBSD-3.1 running on a Pentium-Pro processor. 

The problem is that the Intel performance counters are 40 bits wide and
reading them using the rdmsr instruction results in the higher 24 bits of
the 64 bit value to be undefined (as stated in Intel's manuals). However, the
code in perfmon.c doesn't clear the higher 24 bits after calling rdmsr().

Here's a sample output of the /usr/share/examples/perfmon program (this
particular example montiors the floating point operations in 1 second):
% perfmon -u -e -s 1 -l 1 193
  1: 61572651155456
  total: 61572651155456.000000
  mean: 61572651155456.000000
  clocks (at 166-MHz): 166827310


Here's the sample output after applying a fix to /sys/i386/i386/perfmon.c:
% perfmon -u -e -s 1 -l 1 193
  1: 0
  total: 0.000000
  mean: 0.000000
  clocks (at 166-MHz): 166632751




- Mohit Aron
  aron@cs.rice.edu


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message