Re: [PATCH tip/core/rcu 01/48] rcu: optionally leave lockdep enabled after RCU lockdep splat

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

On Wed, May 05, 2010 at 07:24:57PM -0400, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > On Wed, May 05, 2010 at 06:46:41PM -0400, Mathieu Desnoyers wrote:
> > > * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > > > From: Lai Jiangshan <laijs@cn.fujitsu.com>
> > > > 
> > > > There is no need to disable lockdep after an RCU lockdep splat,
> > > > so remove the debug_lockdeps_off() from lockdep_rcu_dereference().
> > > > To avoid repeated lockdep splats, use a static variable in the inlined
> > > > rcu_dereference_check() and rcu_dereference_protected() macros so that
> > > > a given instance splats only once, but so that multiple instances can
> > > > be detected per boot.
> > > > 
> > > > This is controlled by a new config variable CONFIG_PROVE_RCU_REPEATEDLY,
> > > > which is disabled by default.  This provides the normal lockdep behavior
> > > > by default, but permits people who want to find multiple RCU-lockdep
> > > > splats per boot to easily do so.
> > > 
> > > I'll play the devil's advocate here. (just because that's so much fun)
> > > ;-)
> > > 
> > > If we look at:
> > > 
> > > include/linux/debug_locks.h:
> > > 
> > > static inline int __debug_locks_off(void)
> > > {
> > >         return xchg(&debug_locks, 0);
> > > }
> > > 
> > > We see that all code following a call to "debug_locks_off()" can assume
> > > that it cannot possibly run concurrently with other code following
> > > "debug_locks_off()". Now, I'm not saying that the code we currently have
> > > will necessarily break, but I think it is worth asking if there is any
> > > assumption in lockdep (or RCU lockdep more specifically) about mutual
> > > exclusion after debug_locks_off() ?
> > > 
> > > Because if there is, then this patch is definitely breaking something by
> > > not protecting lockdep against multiple concurrent executions.
> > 
> > So what in lockdep_rcu_dereference() needs to be protected from concurrent
> > execution?  All that happens beyond that point is a bunch of printk()s,
> > printing the locks held by this task, and dumping this task's stack.
> > 
> > 							Thanx, Paul
> 
> I agree with you that printing the current task information should be safe.
> However, I am not sure that concurrent updates to the lock_class while printk()
> is showing its information will end up doing what we expect it to do.
> 
> It could be acceptable to have unreliable information in these rare cases, but
> the important thing would be to ensure that the kernel does not OOPS.

But any races other than the printk()s can already happen as follows:

o	CPU 0 needs to update some information about the lock.  It
	checks debug_locks and finds that it is non-zero.

o	CPU 1 detects a deadlock, and invokes __debug_locks_off(),
	which atomically sets debug_locks to zero.

o	CPU 1 then proceeds to printk() information that CPU 0
	is concurrently modifying.  Which looks to be OK in any case.

Or is there some other race that cannot already happen that I am
introducing?

							Thanx, Paul