Re: rcu, sched: WARNING: CPU: 30 PID: 23771 at kernel/rcu/tree_plugin.h:337 rcu_read_unlock_special+0x369/0x550()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jan 23, 2015 at 05:16:46PM +0800, Lai Jiangshan wrote:
> On 01/23/2015 02:55 PM, Paul E. McKenney wrote:
> > On Thu, Jan 22, 2015 at 11:05:45PM -0500, Sasha Levin wrote:
> >> On 01/22/2015 11:02 PM, Sasha Levin wrote:
> >>> On 01/22/2015 10:51 PM, Paul E. McKenney wrote:
> >>>> On Thu, Jan 22, 2015 at 10:29:01PM -0500, Sasha Levin wrote:
> >>>>>> On 01/21/2015 07:43 PM, Paul E. McKenney wrote:
> >>>>>>>> On Wed, Jan 21, 2015 at 10:44:57AM -0500, Sasha Levin wrote:
> >>>>>>>>>> On 01/20/2015 09:57 PM, Paul E. McKenney wrote:
> >>>>>>>>>>>>>>>> So RCU believes that an RCU read-side critical section that ended within
> >>>>>>>>>>>>>>>>>>>> an interrupt handler (in this case, an hrtimer) somehow got preempted.
> >>>>>>>>>>>>>>>>>>>> Which is not supposed to happen.
> >>>>>>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>>>>> Do you have CONFIG_PROVE_RCU enabled?  If not, could you please enable it
> >>>>>>>>>>>>>>>>>>>> and retry?
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> I did have CONFIG_PROVE_RCU, and didn't see anything else besides what I pasted here.
> >>>>>>>>>>>> OK, fair enough.  I do have a stack of RCU CPU stall-warning changes on
> >>>>>>>>>>>> their way in, please see v3.19-rc1..630181c4a915 in -rcu, which is at:
> >>>>>>>>>>>>
> >>>>>>>>>>>> git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git
> >>>>>>>>>>>>
> >>>>>>>>>>>> These handle the problems that Dave Jones, yourself, and a few others
> >>>>>>>>>>>> located this past December.  Could you please give them a spin?
> >>>>>>>>>>
> >>>>>>>>>> They seem to be a part of -next already, so this testing already includes them.
> >>>>>>>>>>
> >>>>>>>>>> I seem to be getting them about once a day, anything I can add to debug it?
> >>>>>>>>
> >>>>>>>> Could you please try reproducing with the following patch?
> >>>>>>
> >>>>>> Yes, and I've got mixed results. It reproduced, and all I got was:
> >>>>>>
> >>>>>> [  717.645572] ===============================
> >>>>>> [  717.645572] [ INFO: suspicious RCU usage. ]
> >>>>>> [  717.645572] 3.19.0-rc5-next-20150121-sasha-00064-g3c37e35-dirty #1809 Tainted: G        W
> >>>>>> [  717.645572] -------------------------------
> >>>>>> [  717.645572] kernel/rcu/tree_plugin.h:337 rcu_read_unlock() from irq or softirq with blocking in critical section!!!
> >>>>>> [  717.645572] !
> >>>>>> [  717.645572]
> >>>>>> [  717.645572] other info that might help us debug this:
> >>>>>> [  717.645572]
> >>>>>> [  717.645572]
> >>>>>> [  717.645572] rcu_scheduler_active = 1, debug_locks = 1
> >>>>>> [  717.645572] 3 locks held by trinity-c29/16497:
> >>>>>> [  717.645572]  #0:  (&sb->s_type->i_mutex_key){+.+.+.}, at: [<ffffffff81bec373>] lookup_slow+0xd3/0x420
> >>>>>> [  717.645572]  #1:
> >>>>>> [hang]
> >>>>>>
> >>>>>> So the rest of the locks/stack trace didn't get printed, nor the pr_alert() which
> >>>>>> should follow that.
> >>>>>>
> >>>>>> I've removed the lockdep call and will re-run it.
> >>>> Thank you!  You are keeping the pr_alert(), correct?
> >>>
> >>> Yup, just the lockdep call goes away.
> >>
> >> Okay, this reproduced faster than I anticipated:
> >>
> >> [  786.160131] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.239513] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.240503] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.242575] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.243565] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.243565] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.243565] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.243565] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >> [  786.243565] ->rcu_read_unlock_special: 0x100 (b: 0, nq: 1)
> >>
> >> It seems like the WARN_ON_ONCE was hiding the fact it actually got hit couple
> >> of times in a very short interval. Maybe that would also explain lockdep crapping
> >> itself.
> > 
> > OK, that was what I thought was the situation.  I have not yet fully
> > worked out how RCU gets into that state, but in the meantime, here
> > is a patch that should prevent the splats.  (It requires a subtle
> > interaction of quiescent-state detection and the scheduling-clock
> > interrupt.)
> > 
> > 							Thanx, Paul
> > 
> > ------------------------------------------------------------------------
> > 
> > rcu: Clear need_qs flag to prevent splat
> > 
> > If the scheduling-clock interrupt sets the current tasks need_qs flag,
> > but if the current CPU passes through a quiescent state in the meantime,
> > then rcu_preempt_qs() will fail to clear the need_qs flag, which can fool
> > RCU into thinking that additional rcu_read_unlock_special() processing
> > is needed.  This commit therefore clears the need_qs flag before checking
> > for additional processing.
> > 
> > Reported-by: Sasha Levin <sasha.levin@oracle.com>
> > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > 
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index 8669de884445..ec99dc16aa38 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -322,6 +322,7 @@ void rcu_read_unlock_special(struct task_struct *t)
> >  	special = t->rcu_read_unlock_special;
> >  	if (special.b.need_qs) {
> >  		rcu_preempt_qs();
> > +		t->rcu_read_unlock_special.need_qs = false;
> >  		if (!t->rcu_read_unlock_special.s) {
> >  			local_irq_restore(flags);
> >  			return;
> > 
> > .
> 
> 
> rcu_preempt_qs() can be called from rcu_preempt_note_context_switch()
> without irq-disabled. I think it is dangerous, since it touches need_qs and
> passed_quiesce directly and touches rcu_read_unlock_special.b.blocked and
> qs_pending indirectly.  At least it addes burden for me to understand them all.

Yep, disabling interrupts across the call to rcu_preempt_qs() in
rcu_preempt_note_context_switch() would be another way to fix this.
And in fact the code used to disable interrupts across this call.
But I was informed that interrupt disabling in the scheduler fastpath
was socially irresponsible, so I removed it, obviously failing to think
it through.  So the fix above should cover things without re-introducing
the overhead on the scheduler fastpath.

Make sense, or is there some other problem with this?

Here are some that I considered and dismissed as non-problems:

1.	We could take an interrupt just before recording the quiescent
	state, do RCU_SOFTIRQ on return from that interrupt, notice a new
	grace period, and upon return from interrupt to rcu_preempt_qs()
	record a quiescent state.

	But that is OK, because we are in rcu_preempt_qs(), and thus
	are in a quiescent state both before and after the interrupt,
	regardless of what the grace period number might be.

2.	As #1 above, but we get interrupted just after recording
	the quiescent state instead of just before.

	This is also OK.  When the new grace period is noticed, the
	quiescent-state information is cleared.  The current quiescent
	state is ignored, but there will be another quiescent state
	along at some point in the near future.

3.	As #1 above, but after clearing need_qs.  Same result as #2.

Any that I missed?

							Thanx, Paul