On Tue, Oct 13, 2020 at 12:34:06PM +0200, Peter Zijlstra wrote:
> On Mon, Oct 12, 2020 at 02:28:12PM -0700, Paul E. McKenney wrote:
> > It is certainly an accident waiting to happen. Would something like
> > the following make sense?
>
> Sadly no.
>
> > ------------------------------------------------------------------------
> >
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index bfd38f2..52a63bc 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -4067,6 +4067,7 @@ void rcu_cpu_starting(unsigned int cpu)
> >
> > rnp = rdp->mynode;
> > mask = rdp->grpmask;
> > + lockdep_off();
> > raw_spin_lock_irqsave_rcu_node(rnp, flags);
> > WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
> > newcpu = !(rnp->expmaskinitnext & mask);
> > @@ -4086,6 +4087,7 @@ void rcu_cpu_starting(unsigned int cpu)
> > } else {
> > raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > }
> > + lockdep_on();
> > smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
> > }
>
> This will just shut it up, but will not fix the actual problem of that
> spin-lock ending up in trace_lock_acquire() which relies on RCU which
> isn't looking.
>
> What we need here is to supress tracing not lockdep. Let me consider.
We appear to have a similar problem with rcu_report_dead(), it's
raw_spin_unlock()s can end up in trace_lock_release() while we just
killed RCU.
