Re: [PATCH 3/7] lockdep: lock contention tracking

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 2007-05-23 at 12:11 -0400, Jason Baron wrote:
> On Wed, 23 May 2007, Peter Zijlstra wrote:
> 
> > On Wed, 2007-05-23 at 10:40 -0400, Jason Baron wrote:
> > > On Wed, 23 May 2007, Peter Zijlstra wrote:
> > > 
> > > > Count lock contention events per lock class. Additionally track the first four
> > > > callsites that resulted in the contention.
> > > > 
> > > 
> > > I think that we need the total number of locking calls, not just the total 
> > > number of *contended* locking calls, in order to make the data 
> > > significant. Same for waittime. Yes, this pollutes the fastpath. In the 
> > > contention case, its one extra addition, and for the waittime, its a call 
> > > the sched_clock(). I don't think that's too much to pay for much more 
> > > meaningful data.
> > 
> > The holdtime statistics add these numbers.
> > 
> 
> ok, i see what you are saying...however, the 'waittime' statistic as 
> implemented, is only recorded when we don't get the lock immediately. 
> Thus, it's really measuring the waittime when there is lock contention. I 
> understand that in the non-contended case we are only talking a a few 
> cycles, but the fact that the non-contended case is not counted as another 
> waittime of even zero length (so no measurement is required), might skew 
> the stats a bit. For examples, if there was a lock that was almost never 
> contended but one time happened to be contended for a long time, its 
> average wait time would look really high.

I'm not seeing how or why this is a problem. The number of contentions
is reported on the same line, so it should be obvious.

Your definition of wait-time is also obtainable from the numbers, one
would just divide waittime-total by acquisitions, instead of
contentions.