Re: scheduler problems in -next (was: Re: [PATCH 6.4 000/227] 6.4.7-rc1 review)

[Paul E. McKenney]

On Tue, Aug 01, 2023 at 12:11:04PM -0700, Paul E. McKenney wrote:
> On Tue, Aug 01, 2023 at 10:32:45AM -0700, Guenter Roeck wrote:
> > On 7/31/23 14:15, Peter Zijlstra wrote:
> > > On Mon, Jul 31, 2023 at 09:34:29AM -0700, Guenter Roeck wrote:
> > > > > Ha!, I was poking around the same thing. My hack below seems to (so far,
> > > > > <20 boots) help things.
> > > > > 
> > > > 
> > > > So, dumb question:
> > > > How comes this bisects to "sched/fair: Remove sched_feat(START_DEBIT)" ?
> > > 
> > > That commit changes the timings of things; dumb luck otherwise.
> > 
> > Kind of scary. So I only experienced the problem because the START_DEBIT patch
> > happened to be queued roughly at the same time, and it might otherwise have
> > found its way unnoticed into the upstream kernel.

And just to set the record straight, this bug has been in mainline for
about a year, since v5.19.

							Thanx, Paul

> >                                                   That makes me wonder if this
> > or other similar patches may uncover similar problems elsewhere in the kernel
> > (i.e., either hide new or existing race conditions or expose existing ones).
> > 
> > This in turn makes me wonder if it would be possible to define a test which
> > would uncover such problems without the START_DEBIT patch. Any idea ?
> 
> Thank you all for tracking this down!
> 
> One way is to put a schedule_timeout_idle(100) right before the call to
> rcu_tasks_one_gp() from synchronize_rcu_tasks_generic().  That is quite
> specific to this particular issue, but it does have the virtue of making
> it actually happen in my testing.
> 
> There have been a few academic projects that inject delays at points
> chosen by various heuristics plus some randomness.  But this would be
> a bit of a challenge to those because each kernel only passes through
> this window once at boot time.
> 
> Please see below for my preferred fix.  Does this work for you guys?
> 
> Back to figuring out why recent kernels occasionally to blow up all
> rcutorture guest OSes...
> 
> 							Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> index 7294be62727b..2d5b8385c357 100644
> --- a/kernel/rcu/tasks.h
> +++ b/kernel/rcu/tasks.h
> @@ -570,10 +570,12 @@ static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
>  	if (unlikely(midboot)) {
>  		needgpcb = 0x2;
>  	} else {
> +		mutex_unlock(&rtp->tasks_gp_mutex);
>  		set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
>  		rcuwait_wait_event(&rtp->cbs_wait,
>  				   (needgpcb = rcu_tasks_need_gpcb(rtp)),
>  				   TASK_IDLE);
> +		mutex_lock(&rtp->tasks_gp_mutex);
>  	}
>  
>  	if (needgpcb & 0x2) {