Re: [PATCH 2.6.12-rc4] cpuset exit NULL dereference fix

[Robin Holt <holt@sgi.com>]

On Thu, May 26, 2005 at 11:00:10AM +0200, Simon Derr wrote:
> 
> I'm a bit concerned about this. Since there might well be 
> 'notify_on_release' cpusets all over the system, and that there is only 
> one cpuset_sem semaphore, I feel like this 'scaling problem' still exists 
> even with:
> 
> if (notify_on_release(cs)) {
> 	down(&cpuset_sem);
> 	...
> 
> Maybe adding more per-cpuset data such as a per-cpuset removal_sem might 
> be worth it ?

Why not change the atomic into a lock and a refcount.  Grab the lock before
each increment/decrement of the refcount and only continue with the removal
code when the refcount reaches 0.  For a normal cpuset, the refcount could
be biased to 1.  Then child cpusets are created, they could increment their
parent cpuset's refcount.  When the notify_on_release flag used to be set,
we decrement the refcount by one.  Whenever the refcount reaches 0, we
automatically remove the cpuset.  Seems really clear, but would require
touching a larger chunk of the code.

Robin