Re: [RFC] Deadlock via recursive wakeup via RCU with threadirqs

[Paul E. McKenney]

On Fri, Jun 28, 2019 at 05:24:38PM +0900, Byungchul Park wrote:
> On Fri, Jun 28, 2019 at 05:14:32PM +0900, Byungchul Park wrote:
> > On Fri, Jun 28, 2019 at 04:43:50PM +0900, Byungchul Park wrote:
> > > On Fri, Jun 28, 2019 at 04:31:38PM +0900, Byungchul Park wrote:
> > > > On Thu, Jun 27, 2019 at 01:36:12PM -0700, Paul E. McKenney wrote:
> > > > > On Thu, Jun 27, 2019 at 03:17:27PM -0500, Scott Wood wrote:
> > > > > > On Thu, 2019-06-27 at 11:41 -0700, Paul E. McKenney wrote:
> > > > > > > On Thu, Jun 27, 2019 at 02:16:38PM -0400, Joel Fernandes wrote:
> > > > > > > > 
> > > > > > > > I think the fix should be to prevent the wake-up not based on whether we
> > > > > > > > are
> > > > > > > > in hard/soft-interrupt mode but that we are doing the rcu_read_unlock()
> > > > > > > > from
> > > > > > > > a scheduler path (if we can detect that)
> > > > > > > 
> > > > > > > Or just don't do the wakeup at all, if it comes to that.  I don't know
> > > > > > > of any way to determine whether rcu_read_unlock() is being called from
> > > > > > > the scheduler, but it has been some time since I asked Peter Zijlstra
> > > > > > > about that.
> > > > > > > 
> > > > > > > Of course, unconditionally refusing to do the wakeup might not be happy
> > > > > > > thing for NO_HZ_FULL kernels that don't implement IRQ work.
> > > > > > 
> > > > > > Couldn't smp_send_reschedule() be used instead?
> > > > > 
> > > > > Good point.  If current -rcu doesn't fix things for Sebastian's case,
> > > > > that would be well worth looking at.  But there must be some reason
> > > > > why Peter Zijlstra didn't suggest it when he instead suggested using
> > > > > the IRQ work approach.
> > > > > 
> > > > > Peter, thoughts?
> > > > 
> > > 
> > > +cc kernel-team@xxxxxxx
> > > (I'm sorry for more noise on the thread.)
> > > 
> > > > Hello,
> > > > 
> > > > Isn't the following scenario possible?
> > > > 
> > > > The original code
> > > > -----------------
> > > > rcu_read_lock();
> > > > ...
> > > > /* Experdite */
> > > > WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, true);
> > > > ...
> > > > __rcu_read_unlock();
> > > > 	if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
> > > > 		rcu_read_unlock_special(t);
> > > > 			WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > > 			rcu_preempt_deferred_qs_irqrestore(t, flags);
> > > > 		barrier();  /* ->rcu_read_unlock_special load before assign */
> > > > 		t->rcu_read_lock_nesting = 0;
> > > > 
> > > > The reordered code by machine
> > > > -----------------------------
> > > > rcu_read_lock();
> > > > ...
> > > > /* Experdite */
> > > > WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, true);
> > > > ...
> > > > __rcu_read_unlock();
> > > > 	if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
> > > > 		rcu_read_unlock_special(t);
> > > > 		t->rcu_read_lock_nesting = 0; <--- LOOK AT THIS!!!
> > > > 			WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > > 			rcu_preempt_deferred_qs_irqrestore(t, flags);
> > > > 		barrier();  /* ->rcu_read_unlock_special load before assign */
> > > > 
> > > > An interrupt happens
> > > > --------------------
> > > > rcu_read_lock();
> > > > ...
> > > > /* Experdite */
> > > > WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, true);
> > > > ...
> > > > __rcu_read_unlock();
> > > > 	if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
> > > > 		rcu_read_unlock_special(t);
> > > > 		t->rcu_read_lock_nesting = 0; <--- LOOK AT THIS!!!
> > > > <--- Handle an (any) irq
> > > > 	rcu_read_lock();
> > > > 	/* This call should be skipped */
> > > > 	rcu_read_unlock_special(t);
> > > > 			WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > > 			rcu_preempt_deferred_qs_irqrestore(t, flags);
> > > > 		barrier();  /* ->rcu_read_unlock_special load before assign */
> > > > 
> > > > We don't have to handle the special thing twice like this which is one
> > > > reason to cause the problem even though another problem is of course to
> > > > call ttwu w/o being aware it's within a context holding pi lock.
> > > > 
> > > > Apart from the discussion about how to avoid ttwu in an improper
> > > > condition, I think the following is necessary. I may have something
> > > > missing. It would be appreciated if you let me know in case I'm wrong.
> > > > 
> > > > Anyway, logically I think we should prevent reordering between
> > > > t->rcu_read_lock_nesting and t->rcu_read_unlock_special.b.exp_hint not
> > > > only by compiler but also by machine like the below.
> > > > 
> > > > Do I miss something?
> > > > 
> > > > Thanks,
> > > > Byungchul
> > > > 
> > > > ---8<---
> > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > index 3c8444e..9b137f1 100644
> > > > --- a/kernel/rcu/tree_plugin.h
> > > > +++ b/kernel/rcu/tree_plugin.h
> > > > @@ -412,7 +412,13 @@ void __rcu_read_unlock(void)
> > > >  		barrier();  /* assign before ->rcu_read_unlock_special load */
> > > >  		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
> > > >  			rcu_read_unlock_special(t);
> > > > -		barrier();  /* ->rcu_read_unlock_special load before assign */
> > > > +		/*
> > > > +		 * Prevent reordering between clearing
> > > > +		 * t->rcu_reak_unlock_special in
> > > > +		 * rcu_read_unlock_special() and the following
> > > > +		 * assignment to t->rcu_read_lock_nesting.
> > > > +		 */
> > > > +		smp_wmb();
> > 
> > Ah. But the problem is this makes rcu_read_unlock() heavier, which is
> > too bad. Need to consider something else. But I'm still curious about
> > if the scenario I told you is correct?
> 
> Instead, this patch should be replaced with the following:
> 
> ---8<---
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index 3c8444e..f103e98 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -624,8 +624,15 @@ static void rcu_read_unlock_special(struct task_struct *t)
>  
>  	local_irq_save(flags);
>  	irqs_were_disabled = irqs_disabled_flags(flags);
> +
> +	WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> +	/*
> +	 * Prevent reordering between rcu_read_unlock_special.b.exp_hint
> +	 * above and rcu_read_lock_nesting outside of this function.
> +	 */
> +	smp_wmb();

Except that these are manipulated by the current CPU (aside from debug
code), so inter-CPU ordering is not needed.  Plus .exp_hint is just a
heuristic.

Or am I missing something subtle here?  If so, please provide a
step-by-step explanation of the failure scenario.

							Thanx, Paul

> +
>  	if (preempt_bh_were_disabled || irqs_were_disabled) {
> -		WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
>  		/* Need to defer quiescent state until everything is enabled. */
>  		if (irqs_were_disabled) {
>  			/* Enabling irqs does not reschedule, so... */
> @@ -638,7 +645,6 @@ static void rcu_read_unlock_special(struct task_struct *t)
>  		local_irq_restore(flags);
>  		return;
>  	}
> -	WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
>  	rcu_preempt_deferred_qs_irqrestore(t, flags);
>  }
>  
>