Re: [PATCH -next v2] locking/osq_lock: annotate a data race in osq_lock

["Paul E. McKenney" <paulmck@kernel.org>]

On Mon, May 11, 2020 at 06:34:13PM +0100, Will Deacon wrote:
> On Mon, May 11, 2020 at 10:29:18AM -0700, Paul E. McKenney wrote:
> > On Mon, May 11, 2020 at 05:52:17PM +0100, Will Deacon wrote:
> > > On Mon, May 11, 2020 at 09:43:19AM -0700, Paul E. McKenney wrote:
> > > > On Mon, May 11, 2020 at 04:58:13PM +0100, Will Deacon wrote:
> > > > > On Sat, May 09, 2020 at 02:36:54PM -0700, Paul E. McKenney wrote:
> > > > > > diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
> > > > > > index 1f77349..1de006e 100644
> > > > > > --- a/kernel/locking/osq_lock.c
> > > > > > +++ b/kernel/locking/osq_lock.c
> > > > > > @@ -154,7 +154,11 @@ bool osq_lock(struct optimistic_spin_queue *lock)
> > > > > >  	 */
> > > > > >  
> > > > > >  	for (;;) {
> > > > > > -		if (prev->next == node &&
> > > > > > +		/*
> > > > > > +		 * cpu_relax() below implies a compiler barrier which would
> > > > > > +		 * prevent this comparison being optimized away.
> > > > > > +		 */
> > > > > > +		if (data_race(prev->next) == node &&
> > > > > >  		    cmpxchg(&prev->next, node, NULL) == node)
> > > > > >  			break;
> > > > > 
> > > > > I'm fine with the data_race() placement, but I don't find the comment
> > > > > very helpful. We assign the result of a READ_ONCE() to 'prev' in the
> > > > > loop, so I don't think that the cpu_relax() is really relevant.
> > > > 
> > > > Suppose that the compiler loaded a value that was not equal to "node".
> > > > In that case, the cmpxchg() won't happen, so something else must force
> > > > the compiler to do the reload in order to avoid an infinite loop, right?
> > > > Or am I missing something here?
> > > 
> > > Then we just go round the loop and reload prev:
> > > 
> > > 	prev = READ_ONCE(node->prev);
> > > 
> > > which should be enough to stop the compiler, no?
> > 
> > Yes, that would also work.  Either have the cpu_relax() or a barrier()
> > or whatever on the one hand, or, as you say, turn the data_race() into
> > a READ_ONCE().  I personally prefer the READ_ONCE() myself, unless that
> > would undesirably suppress other KCSAN warnings.
> 
> No, I mean here is the code after this patch is applied:
> 
> 	for (;;) {
> 		if (data_race(prev->next) == node &&
> 		    cmpxchg(&prev->next, node, NULL) == node)
> 			break;
> 
> 		/*
> 		 * We can only fail the cmpxchg() racing against an unlock(),
> 		 * in which case we should observe @node->locked becomming
> 		 * true.
> 		 */
> 		if (smp_load_acquire(&node->locked))
> 			return true;
> 
> 		cpu_relax();
> 
> 		/*
> 		 * Or we race against a concurrent unqueue()'s step-B, in which
> 		 * case its step-C will write us a new @node->prev pointer.
> 		 */
> 		prev = READ_ONCE(node->prev);
> 	}
> 
> I'm saying that this READ_ONCE at the end of the loop should be sufficient
> to stop the compiler making value assumptions about prev->next. Do you
> agree?

Good point, and I would certainly hope so!

							Thanx, Paul