On 05/05/21 14:13, Vincent Guittot wrote:

On Thu, 29 Apr 2021 at 01:28, Scott Wood [off-list ref] wrote:

quoted

This is required in order to be able to enable interrupts in the next
patch.  This is limited to PREEMPT_RT to avoid adding potentially
measurable overhead to the non-RT case (requiring a double switch when
pulling a task onto a newly idle cpu).

IIUC, only the newidle_balance is a problem and not the idle load
balance that runs softirq. In this case, why not skipping
newidle_balance entirely in case of preempt_rt and kick an idle load
balance instead as you switch to idle thread context anyway

So if I follow you, that would be along the lines of having PREEMPT_RT turn
newidle_balance() into:

        rq->idle_balance = CPU_IDLE;
        rq->next_balance = jiffies;
        trigger_load_balance(rq);

which I'm thinking isn't too crazy.