Re: [PATCH,RFC] perf: panic due to inclied cpu context task_ctx value

[Oleg Nesterov <oleg@redhat.com>]

On 03/30, Peter Zijlstra wrote:
>
> -atomic_t perf_sched_events __read_mostly;
> +atomic_t perf_sched_events_in __read_mostly;
> +atomic_t perf_sched_events_out __read_mostly;
>  static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
>
> +static void perf_sched_events_inc(void)
> +{
> +	jump_label_inc(&perf_sched_events_out);
> +	jump_label_inc(&perf_sched_events_in);
> +}
> +
> +static void perf_sched_events_dec(void)
> +{
> +	jump_label_dec(&perf_sched_events_in);
> +	JUMP_LABEL(&perf_sched_events_in, no_sync);
> +	synchronize_sched();
> +no_sync:
> +	jump_label_dec(&perf_sched_events_out);
> +}

OK, synchronize_sched() can't work. How about

	static int force_perf_event_task_sched_out(void *unused)
	{
		struct task_struct *curr = current;

		__perf_event_task_sched_out(curr, task_rq(curr)->idle);

		return 0;
	}

	void synchronize_perf_event_task_sched_out(void)
	{
		stop_machine(force_perf_event_task_sched_out, NULL,
				cpu_possible_mask);
	}

instead?

	- stop_machine(cpu_possible_mask) ensures that each cpu
	  does the context switch and calls _sched_out

	- force_perf_event_task_sched_out() is only needed because
	  the migration thread can have the counters too.

Note, I am not sure this is the best solution. Just in case we don't
find something better.

In any case, do you think this can work or I missed something again?

Oleg.