Re: [syzbot] WARNING in exit_tasks_rcu_finish

[Paul E. McKenney]

On Mon, Jun 13, 2022 at 11:54:45PM +0000, Zhang, Qiang1 wrote:
> On Mon, Jun 13, 2022 at 10:26:47PM +0000, Zhang, Qiang1 wrote:
> > On Mon, Jun 13, 2022 at 01:55:31PM +0000, Zhang, Qiang1 wrote:
> > > > syzbot <syzbot+9bb26e7c5e8e4fa7e641@xxxxxxxxxxxxxxxxxxxxxxxxx> writes:
> > > > 
> > > > > Hello,
> > > > >
> > > > > syzbot found the following issue on:
> > > > >
> > > > > HEAD commit:    6d0c80680317 Add linux-next specific files for 20220610
> > > > > git tree:       linux-next
> > > > > console output: 
> > > > > https://syzkaller.appspot.com/x/log.txt?x=13b52c2ff00000
> > > > > kernel config:  
> > > > > https://syzkaller.appspot.com/x/.config?x=a30d6e3e814e5931
> > > > > dashboard link: https://syzkaller.appspot.com/bug?extid=9bb26e7c5e8e4fa7e641
> > > > > compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
> > > > >
> > > > > Unfortunately, I don't have any reproducer for this issue yet.
> > > > 
> > > > I don't understand what is going on in linux-next 
> > > > kernel/rcu/tasks.h looks different than in Linus's tree.  Paul 
> > > > does that mean you have some staged rcu changes?
> > > 
> > > >Less than 100 RCU-related patches in -rcu, so not all that bad.  
> > > >;-)
> > > >
> > > >But yes, this could possibly be an issue in one of those patches.
> > > 
> > > > Eric
> > > > 
> > > > 
> > > > > IMPORTANT: if you fix the issue, please add the following tag to the commit:
> > > > > Reported-by: 
> > > > > syzbot+9bb26e7c5e8e4fa7e641@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > > >
> > > > > ------------[ cut here ]------------
> > > > > WARNING: CPU: 1 PID: 28639 at kernel/rcu/tasks.h:1664 
> > > > > exit_tasks_rcu_finish_trace kernel/rcu/tasks.h:1664 [inline]
> > > > > WARNING: CPU: 1 PID: 28639 at kernel/rcu/tasks.h:1664
> > > > > exit_tasks_rcu_finish+0x122/0x1b0 kernel/rcu/tasks.h:1006
> > > 
> > > >The usual way for this warning to trigger is for these a task to exit while in an RCU Tasks Trace read-side critical section:
> > > >
> > > >	rcu_read_lock_trace();
> > > >	do_something_that_causes_task_exit();
> > > >
> > > 
> > > Hi Paul, wether the following scenarios be considered
> > > 
> > > rcu_read_unlock_trace_special
> > >    ->if (trs.b.blocked)
> > >        ->raw_spin_lock_irqsave_rcu_node
> > >        ->list_del_init(&t->trc_blkd_node)
> > >        ->WRITE_ONCE(t->trc_reader_special.b.blocked, false)
> > >        ->raw_spin_unlock_irqrestore_rcu_node
> > >                      ->Inerrrupt
> > >                             ->schedule
> > >                                 ->rcu_note_context_switch
> > >                                     ->rcu_tasks_trace_qs
> > >                                          If (___rttq_nesting && !READ_ONCE((t)->trc_reader_special.b.blocked) 
> > >                                                         /*___rttq_nesting  ==1    &&  (t)->trc_reader_special.b.blocked =false*/
> > > 				rcu_tasks_trace_qs_blkd(t)     
> > >    ->WRITE_ONCE(t->trc_reader_nesting, 0)
> > >     .......
> > >    -> exit_tasks_rcu_finish
> > > 
> > > Whether the following patch can fix it, or what am I missing?
> > > Any thoughts?
> > > 
> > > diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h index
> > > f1209ce621c5..c607e4c914d3 100644
> > > --- a/kernel/rcu/tasks.h
> > > +++ b/kernel/rcu/tasks.h
> > > @@ -1247,6 +1247,7 @@ void rcu_read_unlock_trace_special(struct task_struct *t)
> > >         struct rcu_tasks_percpu *rtpcp;
> > >         union rcu_special trs;
> > > 
> > > +       WRITE_ONCE(t->trc_reader_nesting, 0);
> > >         // Open-coded full-word version of rcu_ld_need_qs().
> > >         smp_mb(); // Enforce full grace-period ordering.
> > >         trs = smp_load_acquire(&t->trc_reader_special);
> > > @@ -1267,7 +1268,6 @@ void rcu_read_unlock_trace_special(struct task_struct *t)
> > >                 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
> > >                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
> > >         }
> > > -       WRITE_ONCE(t->trc_reader_nesting, 0);
> > >  }
> > >  EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
> > 
> > >Thank you for looking into this!
> > >
> > >You do have what I believe to be the correct failure scenario, but the above fix would break nested RCU Tasks Trace read-side critical sections.
> > 
> > Hi Paul
> > 
> > Break nested RCU Tasks Trace read-side critical sections?  
> > Does it mean the following?
> > 
> > rcu_read_unlock_trace
> >     -> WRITE_ONCE(t->trc_reader_nesting, INT_MIN);
> >          /* t->trc_reader_special.s  == 0*/
> >     -> if (likely(!READ_ONCE(t->trc_reader_special.s)) || nesting)
> >           -> Interrupt
> >               -> schedule
> >                  -> rcu_note_context_switch
> >                      -> rcu_tasks_trace_qs
> >                               /*___rttq_nesting  == INT_MIN    &&  (t)->trc_reader_special.b.blocked == false*/
> >                           ->rcu_tasks_trace_qs_blkd(t)     
> >              /*nesting == 0*/
> >          -> WRITE_ONCE(t->trc_reader_nesting, nesting);
> >          -> return;
> >  .........
> >  exit_tasks_rcu_finish
> >      trigger Warnings
> > 
> > Or where am I misunderstanding?
> 
> >I suspect that you do in fact understand it.  Let's walk through the failure scenario again and see.
> >
> >If that READ_ONCE(t->trc_reader_special.s) return zero as you suggest, and then the interrupt and schedule happen as you suggest, then
> >rcu_tasks_trace_qs_blkd() will see a non-zero ->trc_reader_nesting and a zero .b.blocked.  It queues the task, but rcu_read_unlock_trace() won't check again.  It will set ->trc_reader_nesting to zero and continue.  As you noted, if the task exits in that state, then
> >exit_tasks_rcu_finish_trace() will trigger its WARN_ON_ONCE().
> 
> I understand the problem you describe, Just like the call flow I described above, 
> rcu_read_unlock_trace_special() may not be called.

Very good!

							Thanx, Paul

> Thanks
> Zqiang
>  
> >
> >Your change won't affect this because rcu_read_unlock_trace_special() never gets called.
> >
> >Hence the approach in the patch below.
> >
> >Do you see any failure modes given the below patch?
> 
> > Thanks
> > Zqiang
> > 
> > >
> > >But would you be willing to try out the patch shown below?
> > 
> > I will try to test it.
> 
> >Thank you very much!
> >
> >						Thanx, Paul
> 
> > >---------------------------------------------------------------------
> > >---
> > 
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h index 
> > 08059d8d4f5a7..937a58b3266bf 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -184,7 +184,7 @@ void rcu_tasks_trace_qs_blkd(struct task_struct *t);
> >  		if (likely(!READ_ONCE((t)->trc_reader_special.b.need_qs)) &&	\
> >  		    likely(!___rttq_nesting)) {					\
> >  			rcu_trc_cmpxchg_need_qs((t), 0,	TRC_NEED_QS_CHECKED);	\
> > -		} else if (___rttq_nesting &&					\
> > +		} else if (___rttq_nesting && ___rttq_nesting != INT_MIN &&	\
> >  			   !READ_ONCE((t)->trc_reader_special.b.blocked)) {	\
> >  			rcu_tasks_trace_qs_blkd(t);				\
> >  		}								\
> > diff --git a/include/linux/rcupdate_trace.h 
> > b/include/linux/rcupdate_trace.h index 6f9c358173989..9bc8cbb33340b 
> > 100644
> > --- a/include/linux/rcupdate_trace.h
> > +++ b/include/linux/rcupdate_trace.h
> > @@ -75,7 +75,7 @@ static inline void rcu_read_unlock_trace(void)
> >  	nesting = READ_ONCE(t->trc_reader_nesting) - 1;
> >  	barrier(); // Critical section before disabling.
> >  	// Disable IPI-based setting of .need_qs.
> > -	WRITE_ONCE(t->trc_reader_nesting, INT_MIN);
> > +	WRITE_ONCE(t->trc_reader_nesting, INT_MIN + nesting);
> >  	if (likely(!READ_ONCE(t->trc_reader_special.s)) || nesting) {
> >  		WRITE_ONCE(t->trc_reader_nesting, nesting);
> >  		return;  // We assume shallow reader nesting.