Re: [PATCH -tip/core/rcu 1/6] Cleanups and fixes for RCU in face of heavy CPU-hotplug stress

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Aug 21, 2009 at 04:44:16PM +0200, Ingo Molnar wrote:
> 
> * Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > On Fri, 21 Aug 2009, Ingo Molnar wrote:
> > 
> > > * Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> wrote:
> > > 
> > > > I would not trust this architecture for synchronization tests. 
> > > > There has been reports of a hardware bug affecting the cmpxchg 
> > > > instruction in the field. The load fence normally implied by 
> > > > the semantic seems to be missing. AFAIK, AMD never 
> > > > acknowledged the problem.
> > > 
> > > If cmpxchg was broken i'd be having far worse problems and very 
> > > widely so.
> > 
> > I believe Mathieu is suggesting that the hardware bug is not that 
> > the compare and exchange does not work in cmpxchg, but that it 
> > does not provide an explicit memory barrier. Such a bug is very 
> > hard to trigger, since it requires a race that allows a memory 
> > write/read to cross the cmpxchg, and then have this be in such a 
> > place that it will cause harm.
> 
> We can argue all sorts of exotic hardware bugs really, proof is 
> still needed.
> 
> [...]
> > > That's not a proof of course (it's near impossible to prove the 
> > > lack of a bug), but it's sure a strong indicator and you'll need 
> > > to provide far more proof of misbehavior before i discount a 
> > > bona fide regression on this box.
> > 
> > But with the above said, I totally agree with your point. More 
> > proof must be given before we can discount that another bug 
> > exists.
> 
> Yeah. Especially given that this code was changed recently ;-)

Please rest assured that I consider my code in this area to be suspect.  

							Thanx, Paul