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

[Guenter Roeck]

On 7/31/23 09:14, Peter Zijlstra wrote:
> On Mon, Jul 31, 2023 at 05:08:29PM +0100, Roy Hopkins wrote:
> > On Mon, 2023-07-31 at 16:52 +0200, Peter Zijlstra wrote:
> > > On Mon, Jul 31, 2023 at 07:48:19AM -0700, Guenter Roeck wrote:
> > >
> > > > > I've taken your config above, and the rootfs.ext2 and run-sh from x86/.
> > > > > I've then modified run-sh to use:
> > > > >
> > > > >     qemu-system-x86_64 -enable-kvm -cpu host
> > > > >
> > > > > What I'm seeing is that some boots get stuck at:
> > > > >
> > > > > [    0.608230] Running RCU-tasks wait API self tests
> > > > >
> > > > > Is this the right 'problem' ?
> > > >
> > > >
> > > > Yes, exactly.
> > >
> > > Excellent! Let me prod that with something sharp, see what comes
> > > creeping out.
> >
> > In an effort to get up to speed with this area of the kernel, I've been playing
> > around with this too today and managed to reproduce the problem using the same
> > configuration. I'm completely new to this code but I think I may have found the
> > root of the problem.
> >
> > What I've found is that there is a race condition between starting the RCU tasks
> > grace-period thread in rcu_spawn_tasks_kthread_generic() and a subsequent call
> > to synchronize_rcu_tasks_generic(). This results in rtp->tasks_gp_mutex being
> > locked in the initial thread which subsequently blocks the newly started grace-
> > period thread.
> >
> > The problem is that although synchronize_rcu_tasks_generic() checks to see if
> > the grace-period kthread is running, it uses rtp->kthread_ptr to achieve this.
> > This is only set in the thread entry point and not when the thread is created,
> > meaning that it is set only after the creating thread yields or is preempted. If
> > this has not happened before the next call to synchronize_rcu_tasks_generic()
> > then a deadlock occurs.
> >
> > I've created a debug patch that introduces a new flag in rcu_tasks that is set
> > when the kthread is created and used this in synchronize_rcu_tasks_generic() in
> > place of READ_ONCE(rtp->kthread_ptr). This fixes the issue in my test
> > environment.
> >
> > I'm happy to have a go at submitting a patch for this if it helps.
>
> 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)" ?

Thanks,
Guenter

> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> index 56c470a489c8..b083b5a30025 100644
> --- a/kernel/rcu/tasks.h
> +++ b/kernel/rcu/tasks.h
> @@ -652,7 +658,11 @@ static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
>   	t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
>   	if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
>   		return;
> -	smp_mb(); /* Ensure others see full kthread. */
> +	for (;;) {
> +		cond_resched();
> +		if (smp_load_acquire(&rtp->kthread_ptr))
> +			break;
> +	}
>   }
>   
>   #ifndef CONFIG_TINY_RCU