Date:	Mon, 28 Mar 2016 06:08:41 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Peter Zijlstra <peterz@...radead.org>
Cc:	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
	"Chatre, Reinette" <reinette.chatre@...el.com>,
	Jacob Pan <jacob.jun.pan@...ux.intel.com>,
	Josh Triplett <josh@...htriplett.org>,
	Ross Green <rgkernel@...il.com>,
	John Stultz <john.stultz@...aro.org>,
	Thomas Gleixner <tglx@...utronix.de>,
	lkml <linux-kernel@...r.kernel.org>,
	Ingo Molnar <mingo@...nel.org>,
	Lai Jiangshan <jiangshanlai@...il.com>, dipankar@...ibm.com,
	Andrew Morton <akpm@...ux-foundation.org>,
	rostedt <rostedt@...dmis.org>,
	David Howells <dhowells@...hat.com>,
	Eric Dumazet <edumazet@...gle.com>,
	Darren Hart <dvhart@...ux.intel.com>,
	Frédéric Weisbecker <fweisbec@...il.com>,
	Oleg Nesterov <oleg@...hat.com>,
	pranith kumar <bobby.prani@...il.com>
Subject: Re: rcu_preempt self-detected stall on CPU from 4.5-rc3, since 3.17

On Mon, Mar 28, 2016 at 08:25:47AM +0200, Peter Zijlstra wrote:
> On Sun, Mar 27, 2016 at 02:06:41PM -0700, Paul E. McKenney wrote:
> 
> > > But, you need hotplug for this to happen, right?
> > 
> > I do, but Ross Green is seeing something that looks similar, and without
> > CPU hotplug.
> 
> Yes, but that's two differences so far, you need hotplug and he's on ARM
> (which doesn't have TIF_POLLING_NR).
> 
> So either we're all looking at the wrong thing or these really are two
> different issues.

Given that this failure has grown more probable over the past several
releases, it does seem quite likely that we have more than one bug.
Or maybe a few bugs and additional innocent-bystander commits that make
one or more of the bugs more probable.

> > > We should not be migrating towards, or waking on, CPUs no longer present
> > > in cpu_active_map, and there is a rcu/sched_sync() after clearing that
> > > bit. Furthermore, migration_call() does a sched_ttwu_pending() (waking
> > > any remaining stragglers) before we migrate all runnable tasks off the
> > > dying CPU.
> > 
> > OK, so I should instrument migration_call() if I get the repro rate up?
> 
> Can do, maybe try the below first. (yes I know how long it all takes :/)

OK, will run this today, then run calibration for last night's run this
evening.

Speaking of which, last night's run (disabling TIF_POLLING_NRFLAG)
consisted of 24 two-hour runs.  Six of them had hard hangs, and another
had a hang that eventually unhung of its own accord.  I believe that this
is significantly fewer failures than from a stock kernel, but I could
be wrong, and it will take some serious testing to give statistical
confidence for whatever conclusion is correct.

> > > The other interesting case would be resched_cpu(), which uses
> > > set_nr_and_not_polling() to kick a remote cpu to call schedule(). It
> > > atomically sets TIF_NEED_RESCHED and returns if TIF_POLLING_NRFLAG was
> > > not set. If indeed not, it will send an IPI.
> > > 
> > > This assumes the idle 'exit' path will do the same as the IPI does; and
> > > if you look at cpu_idle_loop() it does indeed do both
> > > preempt_fold_need_resched() and sched_ttwu_pending().
> > > 
> > > Note that one cannot rely on irq_enter()/irq_exit() being called for the
> > > scheduler IPI.
> > 
> > OK, thank you for the info!  Any specific debug actions?
> 
> Dunno, something like the below should bring visibility into the
> (lockless) wake_list thingy.
> 
> So these trace_printk()s should happen between trace_sched_waking() and
> trace_sched_wakeup() (I've not fully read the thread, but ISTR you had
> some traces with these here thingies on).
> 
> ---
>  arch/x86/include/asm/bitops.h | 6 ++++--
>  kernel/sched/core.c           | 9 +++++++++
>  2 files changed, 13 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/include/asm/bitops.h b/arch/x86/include/asm/bitops.h
> index 7766d1cf096e..5345784d5e41 100644
> --- a/arch/x86/include/asm/bitops.h
> +++ b/arch/x86/include/asm/bitops.h
> @@ -112,11 +112,13 @@ clear_bit(long nr, volatile unsigned long *addr)
>  	if (IS_IMMEDIATE(nr)) {
>  		asm volatile(LOCK_PREFIX "andb %1,%0"
>  			: CONST_MASK_ADDR(nr, addr)
> -			: "iq" ((u8)~CONST_MASK(nr)));
> +			: "iq" ((u8)~CONST_MASK(nr))
> +			: "memory");
>  	} else {
>  		asm volatile(LOCK_PREFIX "btr %1,%0"
>  			: BITOP_ADDR(addr)
> -			: "Ir" (nr));
> +			: "Ir" (nr)
> +			: "memory");
>  	}
>  }

Is the above addition of "memory" strictly for the debug below, or is
it also a potential fix?

Starting it up regardless, but figured I should ask!

							Thanx, Paul

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 0b21e7a724e1..b446f73c530d 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1669,6 +1669,7 @@ void sched_ttwu_pending(void)
>  	while (llist) {
>  		p = llist_entry(llist, struct task_struct, wake_entry);
>  		llist = llist_next(llist);
> +		trace_printk("waking %d\n", p->pid);
>  		ttwu_do_activate(rq, p, 0);
>  	}
> 
> @@ -1719,6 +1720,7 @@ static void ttwu_queue_remote(struct task_struct *p, int cpu)
>  	struct rq *rq = cpu_rq(cpu);
> 
>  	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
> +		trace_printk("queued %d for waking on %d\n", p->pid, cpu);
>  		if (!set_nr_if_polling(rq->idle))
>  			smp_send_reschedule(cpu);
>  		else
> @@ -5397,10 +5399,17 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
>  		migrate_tasks(rq);
>  		BUG_ON(rq->nr_running != 1); /* the migration thread */
>  		raw_spin_unlock_irqrestore(&rq->lock, flags);
> +
> +		/* really bad m'kay */
> +		WARN_ON(!llist_empty(&rq->wake_list));
> +
>  		break;
> 
>  	case CPU_DEAD:
>  		calc_load_migrate(rq);
> +
> +		/* more bad */
> +		WARN_ON(!llist_empty(&rq->wake_list));
>  		break;
>  #endif
>  	}
> 

Powered by blists - more mailing lists