Re: [PATCH 02/11] rcu: fix bug when rcu_exp_handler() in nested interrupt

[Paul E. McKenney]

On Thu, Oct 31, 2019 at 06:47:31AM -0700, Paul E. McKenney wrote:
> On Thu, Oct 31, 2019 at 10:07:57AM +0000, Lai Jiangshan wrote:
> > These is a possible bug (although which I can't triger yet)
> > since 2015 8203d6d0ee78
> > (rcu: Use single-stage IPI algorithm for RCU expedited grace period)
> > 
> >  rcu_read_unlock()
> >   ->rcu_read_lock_nesting = -RCU_NEST_BIAS;
> >   interrupt(); // before or after rcu_read_unlock_special()
> >    rcu_read_lock()
> >     fetch some rcu protected pointers
> >     // exp GP starts in other cpu.
> >     some works
> >     NESTED interrupt for rcu_exp_handler();

Also, which platforms support nested interrupts?  Last I knew, this was
prohibited.

> >       report exp qs! BUG!
> 
> Why would a quiescent state for the expedited grace period be reported
> here?  This CPU is still in an RCU read-side critical section, isn't it?

And I now see what you were getting at here.  Yes, the current code
assumes that interrupt-disabled regions, like hardware interrupt
handlers, cannot be interrupted.  But if interrupt-disabled regions such
as hardware interrupt handlers can be interrupted (as opposed to being
NMIed), wouldn't that break a whole lot of stuff all over the place in
the kernel?  So that sounds like an arch bug to me.

							Thanx, Paul

> >     // exp GP completes and pointers are freed in other cpu
> >     some works with the pointers. BUG
> >    rcu_read_unlock();
> >   ->rcu_read_lock_nesting = 0;
> > 
> > Although rcu_sched_clock_irq() can be in nested interrupt,
> > there is no such similar bug since special.b.need_qs
> > can only be set when ->rcu_read_lock_nesting > 0
> > 
> > Signed-off-by: Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx>
> > ---
> >  kernel/rcu/tree_exp.h    | 5 +++--
> >  kernel/rcu/tree_plugin.h | 9 ++++++---
> >  2 files changed, 9 insertions(+), 5 deletions(-)
> > 
> > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > index 6dec21909b30..c0d06bce35ea 100644
> > --- a/kernel/rcu/tree_exp.h
> > +++ b/kernel/rcu/tree_exp.h
> > @@ -664,8 +664,9 @@ static void rcu_exp_handler(void *unused)
> >  	 * Otherwise, force a context switch after the CPU enables everything.
> >  	 */
> >  	rdp->exp_deferred_qs = true;
> > -	if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
> > -	    WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs())) {
> > +	if (rcu_preempt_need_deferred_qs(t) &&
> > +	    (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
> > +	    WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs()))) {
> >  		rcu_preempt_deferred_qs(t);
> >  	} else {
> >  		set_tsk_need_resched(t);
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index d4c482490589..59ef10da1e39 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -549,9 +549,12 @@ rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
> >   */
> >  static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
> >  {
> > -	return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
> > -		READ_ONCE(t->rcu_read_unlock_special.s)) &&
> > -	       t->rcu_read_lock_nesting <= 0;
> > +	return (__this_cpu_read(rcu_data.exp_deferred_qs) &&
> > +		(!t->rcu_read_lock_nesting ||
> > +		 t->rcu_read_lock_nesting == -RCU_NEST_BIAS))
> > +		||
> > +		(READ_ONCE(t->rcu_read_unlock_special.s) &&
> > +		 t->rcu_read_lock_nesting <= 0);
> >  }
> >  
> >  /*
> > -- 
> > 2.20.1
> >