Re: [RFC/RFT PATCH 1/2] sched/core: Check and schedule ksoftirq

[Frederic Weisbecker]

On Fri, Dec 16, 2022 at 12:19:34PM +0100, Peter Zijlstra wrote:
> On Thu, Dec 15, 2022 at 10:42:59AM -0800, Srinivas Pandruvada wrote:
> > +		/* Give ksoftirqd 1 jiffy to get a chance to start its job */
> > +		if (!READ_ONCE(it.done) && task_is_running(__this_cpu_read(ksoftirqd))) {
> > +			__set_current_state(TASK_UNINTERRUPTIBLE);
> > +			schedule_timeout(1);
> > +		}
> 
> That's absolutely disgusting :-/

I know, and I hate checking task_is_running(__this_cpu_read(ksoftirqd))
everywhere in idle. And in fact it doesn't work because some cpuidle drivers
also do need_resched() checks.

I guess that either we assume that the idle injection is more important
than serving softirqs and we shutdown the warnings accordingly, or we arrange
for idle injection to have a lower prio than ksoftirqd.

Thoughts?