On Fri, Aug 17, 2007 at 08:53:57AM -0700, Paul E. McKenney wrote:
> On Fri, Aug 17, 2007 at 09:56:45AM +0200, Peter Zijlstra wrote:
> > On Thu, 2007-08-16 at 09:01 -0700, Paul E. McKenney wrote:
> > > On Thu, Aug 16, 2007 at 04:25:07PM +0200, Peter Zijlstra wrote:
> > > >
> > > > There seem to be some unbalanced rcu_read_{,un}lock() issues of late,
> > > > how about doing something like this:
> > >
> > > This will break when rcu_read_lock() and rcu_read_unlock() are invoked
> > > from NMI/SMI handlers -- the raw_local_irq_save() in lock_acquire() will
> > > not mask NMIs or SMIs.
> > >
> > > One approach would be to check for being in an NMI/SMI handler, and
> > > to avoid calling lock_acquire() and lock_release() in those cases.
> >
> > It seems:
> >
> > #define nmi_enter() do { lockdep_off(); __irq_enter(); } while (0)
> > #define nmi_exit() do { __irq_exit(); lockdep_on(); } while (0)
> >
> > Should make it all work out just fine. (for NMIs at least, /me fully
> > ignorant of the workings of SMIs)
>
> Very good point, at least for NMIs on i386 and x86_64. Can't say that I
> know much about SMIs myself. Or about whatever equivalents to NMIs and
> SMIs might exist on other platforms. :-/ Of course, the other platforms
> could be handled by making the RCU lockdep operate only on i386 and x86_64
> if required.
>
> Corey, any advice on SMI handlers? Is there something like nmi_enter()
> and nmi_exit() that allows disabing lockdep?
>
> > > Another approach would be to use sparse, which has checks for
> > > rcu_read_lock()/rcu_read_unlock() nesting.
> >
> > Yeah, but one more method can never hurt, no? :-)
>
> Excellent point!
>
> I guess the next thing for me is to do a performance check. Looks like
> CONFIG_DEBUG_LOCK_ALLOC is not on by default, so not violently worried
> about performance, but would be good to know.
4-CPU 1.8GHz Opteron 844 running 2.6.22 CONFIG_PREEMPT_NONE with Peter's
patch running rcu_read_lock()/rcu_read_unlock() in a tight loop:
1 CPU: 91.7ns
2 CPU: 335.2ns
3 CPU: 534.3ns (But bimodal: two CPUs at 640.7ns, one at 321.6ns)
4 CPU: 784.0ns (again bimodal: three CPUs at 895.9ns, one at 448.2ns)
Running without Peter's patch measures the loop overhead of about 1.2ns.
So not crippling, but definitely a debug-only option that is not enabled
by default in production.
In summary, seems like an eminently reasonable approach to me, assuming
that the NMI/SMI issues can be resolved across the platforms that have
them. Good stuff, Peter!!!
Thanx, Paul
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
