Re: RCU used on incoming CPU before rcu_cpu_starting() called

[Paul E. McKenney]

On Thu, Mar 09, 2017 at 07:29:26AM -0800, Paul E. McKenney wrote:
> On Thu, Mar 09, 2017 at 04:12:55PM +0100, Peter Zijlstra wrote:
> > On Thu, Mar 09, 2017 at 02:08:23PM +0100, Thomas Gleixner wrote:
> > > On Wed, 8 Mar 2017, Paul E. McKenney wrote:
> > > > [   30.694013]  lockdep_rcu_suspicious+0xe7/0x120
> > > > [   30.694013]  get_work_pool+0x82/0x90
> > > > [   30.694013]  __queue_work+0x70/0x5f0
> > > > [   30.694013]  queue_work_on+0x33/0x70
> > > > [   30.694013]  clear_sched_clock_stable+0x33/0x40
> > > > [   30.694013]  early_init_intel+0xe7/0x2f0
> > > > [   30.694013]  init_intel+0x11/0x350
> > > > [   30.694013]  identify_cpu+0x344/0x5a0
> > > > [   30.694013]  identify_secondary_cpu+0x18/0x80
> > > > [   30.694013]  smp_store_cpu_info+0x39/0x40
> > > > [   30.694013]  start_secondary+0x4e/0x100
> > > > [   30.694013]  start_cpu+0x14/0x14
> > > > 
> > > > Here is the relevant code from x86's smp_callin():
> > > > 
> > > > 	/*
> > > > 	 * Save our processor parameters. Note: this information
> > > > 	 * is needed for clock calibration.
> > > > 	 */
> > > > 	smp_store_cpu_info(cpuid);
> > > >
> > > > The problem is that smp_store_cpu_info() indirectly invokes
> > > > schedule_work(), which wants to use RCU.  But RCU isn't informed
> > > > of the incoming CPU until the call to notify_cpu_starting(), which
> > > > causes lockdep to complain bitterly about the use of RCU by the
> > > > premature call to schedule_work().
> > > 
> > > Right. And that want's to be fixed, not hacked around by silencing RCU.
> > > 
> > > Peter????
> > 
> > I'm thinking this is hotplug? 30 seconds after boot is far too late for
> > SMP bringup, or you have a stupid slow machine.
> 
> And this certainly does qualify as "shortly", thank you!
> 
> Yes, this only happens on hotplug with lockdep enabled, specifically
> on rcutorture scenarios TASKS01 and TREE05.
> 
> > Because it only calls schedule_work() after SMP-init. In which case
> > there's then two cases, either:
> > 
> >  - TSC was stable, hotplug wrecked it, TSC is now unstable, and we're
> >    screwed.
> > 
> >  - TSC was unstable, hotplug triggers and we want to mark it unstable
> >    _again_.
> > 
> > If this is the second, the below should fix it, if its the first, I've
> > no idea yet on how to fix that properly :/
> 
> I have applied this patch and started tests on TREE05 and TASKS01, should
> get results shortly.

And the below patch passed light rcutorture testing, so looking good!

							Thanx, Paul

> > Bloody hotplug..
> > 
> > ---
> >  kernel/sched/clock.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
> > index a08795e..eecf388 100644
> > --- a/kernel/sched/clock.c
> > +++ b/kernel/sched/clock.c
> > @@ -172,7 +172,7 @@ void clear_sched_clock_stable(void)
> > 
> >  	smp_mb(); /* matches sched_clock_init_late() */
> > 
> > -	if (sched_clock_running == 2)
> > +	if (sched_clock_running == 2 && sched_clock_stable())
> >  		schedule_work(&sched_clock_work);
> >  }
> > 
> >