Re: [RFC PATCH 3/6] sched/dl: Try better placement even for deadline tasks that do not block

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jul 11, 2019 at 01:17:17PM +0200, Dietmar Eggemann wrote:
> On 7/9/19 3:42 PM, Peter Zijlstra wrote:

> >>> That is, we only do those callbacks from:
> >>>
> >>>   schedule_tail()
> >>>   __schedule()
> >>>   rt_mutex_setprio()
> >>>   __sched_setscheduler()
> >>>
> >>> and the above looks like it can happen outside of those.

> Is this what you are concerned about?
> 
> (2 Cpus (CPU1, CPU2), 4 deadline task (thread0-X)) with 
> 
> @@ -1137,6 +1137,13 @@ static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
>         rf->cookie = lockdep_pin_lock(&rq->lock);
>  
>  #ifdef CONFIG_SCHED_DEBUG
> +#ifdef CONFIG_SMP
> +       /*
> +        * There should not be pending callbacks at the start of rq_lock();
> +        * all sites that handle them flush them at the end.
> +        */
> +       WARN_ON_ONCE(rq->balance_callback);
> +#endif
> 
> 
> [   87.251237] *** <--- queue_balance_callback(migrate_dl_task) p=[thread0-3 3627] on CPU2
> [   87.251261] WARNING: CPU: 2 PID: 3627 at kernel/sched/sched.h:1145 __schedule+0x56c/0x690
> [   87.615882] WARNING: CPU: 2 PID: 3616 at kernel/sched/sched.h:1145 task_rq_lock+0xe8/0xf0
> [   88.176844] WARNING: CPU: 2 PID: 3616 at kernel/sched/sched.h:1145 load_balance+0x4d0/0xbc0
> [   88.381905] WARNING: CPU: 2 PID: 3616 at kernel/sched/sched.h:1145 load_balance+0x7d8/0xbc0

I'm not sure how we get 4 warns, I was thinking that as soon as we exit
__schedule() we'd procress the callback so further warns would be
avoided.

> [   88.586991] *** ---> migrate_dl_task() p=[thread0-3 3627] to CPU1

But yes, something like this. Basucally I want to avoid calling
queue_balance_callback() from a context where we'll not follow up with
balance_callback().