Re: [RFC PATCH 6/8] sched: Report local wake up on resched blind zone within idle loop

[Peter Zijlstra]

On Sat, Jan 09, 2021 at 03:05:34AM +0100, Frederic Weisbecker wrote:
> The idle loop has several need_resched() checks that make sure we don't
> miss a rescheduling request. This means that any wake up performed on
> the local runqueue after the last generic need_resched() check is going
> to have its rescheduling silently ignored. This has happened in the
> past with rcu kthreads awaken from rcu_idle_enter() for example.
> 
> Perform sanity checks to report these situations.

I really don't like this..

 - it's too specific to the actual reschedule condition, any wakeup this
   late is dodgy, not only those that happen to cause a local
   reschedule.

 - we can already test this with unwind and checking against __cpuidle

 - moving all of __cpuidle into noinstr would also cover this. And we're
   going to have to do that anyway.

> +void noinstr sched_resched_local_assert_allowed(void)
> +{
> +	if (this_rq()->resched_local_allow)
> +		return;
> +

> +	/*
> +	 * Idle interrupts break the CPU from its pause and
> +	 * rescheduling happens on idle loop exit.
> +	 */
> +	if (in_hardirq())
> +		return;
> +
> +	/*
> +	 * What applies to hardirq also applies to softirq as
> +	 * we assume they execute on hardirq tail. Ksoftirqd
> +	 * shouldn't have resched_local_allow == 0.
> +	 * We also assume that no local_bh_enable() call may
> +	 * execute softirqs inline on fragile idle/entry
> +	 * path...
> +	 */
> +	if (in_serving_softirq())
> +		return;
> +
> +	WARN_ONCE(1, "Late current task rescheduling may be lost\n");

That seems like it wants to be:

	WARN_ONCE(in_task(), "...");

> +}