Re: [1/4] rcu: Detect uses of rcu read side in extended quiescent states

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

On Tue, Jun 07, 2011 at 08:49:01PM +0200, Frederic Weisbecker wrote:
> On Tue, Jun 07, 2011 at 11:34:14AM -0700, Paul E. McKenney wrote:
> > On Tue, Jun 07, 2011 at 02:58:13PM +0200, Frederic Weisbecker wrote:
> > > On Mon, Jun 06, 2011 at 09:40:05PM -0700, Paul E. McKenney wrote:> 
> > > > The bit I am missing is how to distinguish between spinlocks (where
> > > > sleeping is illegal) and mutexes (where sleeping is perfectly fine).
> > > > We could teach lockdep the difference, I suppose, but it is not clear
> > > > to me that it is worth it.
> > > 
> > > Ah, in fact it doesn't pass through any lockdep check.
> > > 
> > > It's only a function called might_sleep() that is placed in functions
> > > that can sleep. And inside might_sleep() it checks whether it is in a preemptible
> > > area. So it's actually locking-agnostic, it only relies on the preempt_count
> > > and some more for the preempt rcu cases.
> > > 
> > > I think it is called CONFIG_DEBUG_SPINLOCK_SLEEP because it was first used
> > > for spinlock debugging purposes. But then it has a broader use now: sleep
> > > inside preemptible section, sleep inside interrupts, sleep inside rcu.
> > 
> > But the __might_sleep() function can only differentiate between
> > spinlocks and sleeplocks if CONFIG_PREEMPT=y.
> 
> It doesn't differentiate between locks but checks on the lowest level
> by looking at the preempt count. But yeah it only works if CONFIG_PREEMPT,
> which is why I proposed to inc/dec the preempt count also when we have
> that DEBUG_SPINLOCK_SLEEP.

Ah, I missed your proposal to inc/dec preempt_count for PREEMPT=n
and DEBUG_SPINLOCK_SLEEP=y.

> > > It certainly deserves a rename, like CONFIG_DEBUG_ILLEGAL_SLEEP.
> > 
> > Hmmm...  It already checks for sleeping in the middle of a
> > preempt_disable() as well as in a spinlock critical section.
> > So the need for a rename is independent of any RCU checking.
> 
> Sure, rcu just adds itself to the pile of users of might_sleep(), thus
> it would be a nice cleanup to rename the option to something more
> generic. But that rename is not necessary to improve RCU checking.

Agreed!

> > > > In contrast, with RCU, this is straightforward -- check for rcu_sched
> > > > and rcu_bh, but not SRCU.
> > 
> > Actually it makes sense to keep the checks in rcu_note_context_switch(),
> > as there are places that call schedule() directly without a might_sleep().
> > Perhaps having checks in both places is the correct approach?
> 
> In this case it makes more sense to add your checks in schedule_debug(),
> so that we don't wait for a context switch to detect the bug.

You might well be right -- looking at it.

							Thanx, Paul