On Thu, 13 Apr 2017 10:44:53 +0200
Peter Zijlstra <pet...@infradead.org> wrote:
> On Wed, Apr 12, 2017 at 02:27:44PM -0400, Steven Rostedt wrote:
> > + * schedule_idle() is similar to schedule_preempt_disable() except
> > + * that it never enables preemption.
>
> That's not right. The primary distinction is that it doesn't call
> sched_submit_work().
That has nothing to do with fixing synchronize_rcu_tasks(), which is
the entire point of my patch, thus it is *not* the primary distinction.
Keeping schedule from enabling preemption and calling functions is the
bug fix. Not calling sched_submit_work() is just an added optimization
benefit.
The point of the patch is to stop idle from enabling preemption,
because it doesn't need to, as sched_submit_work() is a nop for it.
I'll update my change log to mention that.
>
> And because that function is a no-op for the idle thread, the idle
> thread can do without calling that and therefore avoid the preemption
> window.
>
> You also need a few words about fake idle threads, search play_idle()
> callers.
Thanks, this is the first I heard of these. I'll go look at them.
>
> You could also make schedule_idle() more robust by adding a WARN for the
> blk_schedule_flush_plug() condition.
Why? The call to schedule_preempt_disabled() never got that far when
coming from do_idle().
static inline void sched_submit_work(struct task_struct *tsk)
{
if (!tsk->state || tsk_is_pi_blocked(tsk))
return;
/*
* If we are going to sleep and we have plugged IO queued,
* make sure to submit it to avoid deadlocks.
*/
if (blk_needs_flush_plug(tsk))
blk_schedule_flush_plug(tsk);
}
Isn't tsk->state always zero for the idle task?
A better case would be WARN_ON(tsk->state)
>
>
> You Changelog is still entirely long and rambling but fails to mention
> the fundamental important stuff :-(
Remember, this patch is to fix a bug and not to optimize idle, although
that is an added benefit. The bug I am fixing, which is in linux-next
now, is that the idle thread breaks synchronize_rcu_tasks() when
calling schedule() with preemption enabled. That's what my ramblings in
the change log are talking about.
-- Steve
