Re: cpu stopper threads and load balancing leads to deadlock

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

On Thu, May 03, 2018 at 06:45:08PM +0200, Peter Zijlstra wrote:
> On Thu, May 03, 2018 at 09:12:31AM -0700, Paul E. McKenney wrote:
> > On Thu, May 03, 2018 at 04:44:50PM +0200, Peter Zijlstra wrote:
> > > On Thu, May 03, 2018 at 04:16:55PM +0200, Mike Galbraith wrote:
> > > > On Thu, 2018-05-03 at 15:56 +0200, Peter Zijlstra wrote:
> > > > > On Thu, May 03, 2018 at 03:32:39PM +0200, Mike Galbraith wrote:
> > > > > 
> > > > > > Dang.  With $subject fix applied as well..
> > > > > 
> > > > > That's a NO then... :-(
> > > > 
> > > > Could say who cares about oddball offline wakeup stat. <cringe>
> > > 
> > > Yeah, nobody.. but I don't want to have to change the wakeup code to
> > > deal with this if at all possible. That'd just add conditions that are
> > > 'always' false, except in this exceedingly rare circumstance.
> > > 
> > > So ideally we manage to tell RCU that it needs to pay attention while
> > > we're doing this here thing, which is what I thought RCU_NONIDLE() was
> > > about.
> > 
> > One straightforward approach would be to provide a arch-specific
> > Kconfig option that tells notify_cpu_starting() not to bother invoking
> > rcu_cpu_starting().  Then x86 selects this Kconfig option and invokes
> > rcu_cpu_starting() itself early enough to avoid splats.
> > 
> > See the (untested, probably does not even build) patch below.
> > 
> > I have no idea where to insert either the "select" or the call to
> > rcu_cpu_starting(), so I left those out.  I know that putting the
> > call too early will cause trouble, but I have no idea what constitutes
> > "too early".  :-/
> 
> Something like so perhaps? Mike, can you play around with that? Could
> burn your granny and eat your cookies.
> 
> 
> diff --git a/arch/x86/kernel/cpu/mtrr/main.c b/arch/x86/kernel/cpu/mtrr/main.c
> index 7468de429087..07360523c3ce 100644
> --- a/arch/x86/kernel/cpu/mtrr/main.c
> +++ b/arch/x86/kernel/cpu/mtrr/main.c
> @@ -793,6 +793,9 @@ void mtrr_ap_init(void)
> 
>  	if (!use_intel() || mtrr_aps_delayed_init)
>  		return;
> +
> +	rcu_cpu_starting(smp_processor_id());
> +
>  	/*
>  	 * Ideally we should hold mtrr_mutex here to avoid mtrr entries
>  	 * changed, but this routine will be called in cpu boot time,
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 2a734692a581..4dab46950fdb 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3775,6 +3775,8 @@ int rcutree_dead_cpu(unsigned int cpu)
>  	return 0;
>  }
> 
> +static DEFINE_PER_CPU(int, rcu_cpu_started);
> +
>  /*
>   * Mark the specified CPU as being online so that subsequent grace periods
>   * (both expedited and normal) will wait on it.  Note that this means that
> @@ -3796,6 +3798,11 @@ void rcu_cpu_starting(unsigned int cpu)
>  	struct rcu_node *rnp;
>  	struct rcu_state *rsp;
> 
> +	if (per_cpu(rcu_cpu_started, cpu))

I would log a non-splat dmesg the first time this happened, just for my
future sanity, but otherwise looks fine.  I am a bit concerned about
calls to rcu_cpu_starting() getting sprinkled all through the code.
Or am I being excessively paranoid?

							Thanx, Paul

> +		return;
> +
> +	per_cpu(rcu_cpu_started, cpu) = 1;
> +
>  	for_each_rcu_flavor(rsp) {
>  		rdp = per_cpu_ptr(rsp->rda, cpu);
>  		rnp = rdp->mynode;
> @@ -3852,6 +3859,8 @@ void rcu_report_dead(unsigned int cpu)
>  	preempt_enable();
>  	for_each_rcu_flavor(rsp)
>  		rcu_cleanup_dying_idle_cpu(cpu, rsp);
> +
> +	per_cpu(rcu_cpu_started, cpu) = 0;
>  }
> 
>  /* Migrate the dead CPU's callbacks to the current CPU. */
>