Re: [RFC][PATCH 3/7] sched: Allow balance callbacks for check_class_changed()

[Kirill Tkhai]

Ð ÐÑ, 03/06/2015 Ð 09:32 +0200, Peter Zijlstra ÐÐÑÐÑ:
> On Tue, Jun 02, 2015 at 07:27:19PM +0300, Kirill Tkhai wrote:
> > > +	 * task_dead_dl() will cancel our timer if we happen to die while
> > > +	 * its still pending.
> > 
> > task_dead_dl() is called for tasks of deadline class only. So if we do that,
> > the timer may be executed after final task's dead.
> 
> Indeed; sleep deprived brain misses the obvious :/
> 
> I can't seem to come up with anything much better than pulling that
> hrtimer_cancel() into finish_task_switch(), however sad that is.

Yeah, class-specific manipulation in generic function finish_task_switch()
worsen modularity.

I have an idea, but it require small hrtimer modification. We may use
task_struct counters {get,put}_task_struct(), when we're starting or cancelling
the timer, and in timer handler. But it require to return back the commit:

commit 61699e13072a89880aa584dcc64c6da465fb2ccc
Author: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Date:   Tue Apr 14 21:09:23 2015 +0000

    hrtimer: Remove hrtimer_start() return value

because restart of the timer races with time handler.

Also, it's not clearly for me if it's safe to do the final put_task_struct()
from irq context and how it worsens life of RT people.

> Something like:
> 
>   for_each_class(class) {
> 	if (class->task_dead)
> 		class->task_dead(prev);
>   }
> 
> Would be nicest, but will slow down the common case. And a callback list
> has the downside of having to preallocate entries for every task,
> causing mostly pointless memory overhead.
> 
> 


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/