Message-ID: <001501c12a6f$fc035da0$cb149aa3@engr.sgi.com>
From: "John Hawkes" <hawkes@sgi.com>
To: <kapish@ureach.com>, <lockmeter@oss.sgi.com>
References: <200108211434.KAA23810@www20.ureach.com>
Subject: Re: query related to spinlock metering
Date: Tue, 21 Aug 2001 11:34:59 -0700
MIME-Version: 1.0
Content-Type: text/plain;
	charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
Sender: owner-lockmeter@oss.sgi.com
Precedence: bulk

From: "Kapish K" <kapish@ureach.com>
>       We have been facing some problems with some kernel defects
> and in the process we felt it might be worthwhile to try some
> wrapper around the spinlock code to see if it might be getting
> called wrongly or released wrongly. So, what we would like to is
> just a wrapper around the spinlock code to do some preliminary
> checks to make sure the lock is held before releasing it or
> along similar lines.

SGI's Lockmeter does not provide any such verification.  Lockmeter
instruments spinlocks in order to report who locks what spinlocks, how
often, for how long, and suffering what wait-times in the event of lock
contention.

> Is this lockmeter from sgi similar in performance and is it
> lightweight?

Lockmeter is lighter-weight than some spinlock instumention
implementations and heavier-weight than other implementations.  Typical
system slowdowns are in the 5-15% range.  And yes, there is potential
for perturbing timing-sensitive locking patterns.

If you want to verify that the lock is held when you do an unlock, then
you need to roll your own wrapper.

John Hawkes
hawkes@sgi.com