Re: [PATCH] sched: fix conflict of schedule domain balance in RT scheduling

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 2011-06-01 at 21:58 +0800, Hillf Danton wrote:

> And checking for cache-hot is also added to confirm the comment there.

Please read that comment.

>  	cpu = task_cpu(p);
> @@ -1253,7 +1253,8 @@ static int find_lowest_rq(struct task_struct *task)
>  	 * We prioritize the last cpu that the task executed on since
>  	 * it is most likely cache-hot in that location.
>  	 */
> -	if (cpumask_test_cpu(cpu, lowest_mask))
> +	if (cpumask_test_cpu(cpu, lowest_mask) &&
> +	    task_hot(task, task_rq(task)->clock_task, NULL))
>  		return cpu;

What task_hot() checks for and what we are assuming are two different
things. In fact, we can disable task_hot() so it always fails. That's
not what we want.

If the task happens to have ran on a CPU that is in the lowest_mask, we
want that CPU. Time may not matter. If we know a task ran on a
particular CPU last, we want to run it there if possible. It may still
have cache lines for it, even if it has been a long time since it last
ran.

-- Steve