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

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 2011-06-03 at 22:30 +0800, Hillf Danton wrote:

> I note the comment related to cache-hot, it is fine for woken task even though
> cache-hot is considered to be nop for RT task,

I would not say cache-hot is a nop for RT tasks. Although RT tasks
should not depend on cache being hot, we don't want to go out of our way
to hurt RT tasks.


>  but what dose it help pushing
> RT tasks waiting on RQ?

What the lowest_mask returns is a mask of all cpus with the lowest
priority. Due to the double_lock_balance() we could have unlocked the rq
lock, and things could have changed. Thus, the current CPU could have
lowered its priority. Unlikely to be the case, but can happen.

If the lowest_mask has the task's CPU set, why push this RT task
somewhere else? Most likely the current CPU will not be set, but lets
not punish the RT task if it is.

-- Steve