update on hrtimer based select/poll and range-hrtimers

update on hrtimer based select/poll and range-hrtimers

[Arjan van de Ven]

Hi,

since the last lkml posting I've merged a few fixes and added comments
from Peter, and I've redone the "estimate_accuracy" function.

Rather than reposting the entire series, I'll point to the git tree at

git://git.kernel.org/pub/scm/linux/kernel/git/arjan/linux-2.6-hrtimer.git

and I've pasted the new function below.
Rather than having the hardcoded steps from the "Linus" function, I've
turned it into:
0 for realtime tasks
"0.1% of the time" for not-nice, not realtime tasks
"0.5% of the time" for nice, not realtime tasks
with a cap of 100msec for both.

I would like to request feedback on this approach; I think this is
better than the "hardcoded steps" as before, but maybe someone can come
up with an ever better idea....

static unsigned long __estimate_accuracy(struct timespec *tv)
{
        unsigned long slack;
        int divfactor = 1000;
  
        if (task_nice(current))
                divfactor = divfactor / 5;

        slack = tv->tv_nsec / divfactor; 
        slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);

        if (slack > 100 * NSEC_PER_MSEC)
                slack =  100 * NSEC_PER_MSEC;
        return slack;
}

static unsigned long estimate_accuracy(struct timespec *tv)
{
        unsigned long ret;
        struct timespec now;

        /*
         * Realtime tasks get a slack of 0 for obvious reasons.
         */

        if (current->policy == SCHED_FIFO ||
                current->policy == SCHED_RR)
                return 0;

        ktime_get_ts(&now);
        now = timespec_sub(*tv, now);
        ret = __estimate_accuracy(&now);
        if (ret < current->timer_slack_ns)
                return current->timer_slack_ns;
        return ret;
}

Re: update on hrtimer based select/poll and range-hrtimers

[Peter Zijlstra]

On Sun, 2008-09-07 at 11:11 -0700, Arjan van de Ven wrote:
> Hi,
> 
> since the last lkml posting I've merged a few fixes and added comments
> from Peter, and I've redone the "estimate_accuracy" function.
> 
> Rather than reposting the entire series, I'll point to the git tree at
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/arjan/linux-2.6-hrtimer.git
> 
> and I've pasted the new function below.
> Rather than having the hardcoded steps from the "Linus" function, I've
> turned it into:
> 0 for realtime tasks
> "0.1% of the time" for not-nice, not realtime tasks
> "0.5% of the time" for nice, not realtime tasks
> with a cap of 100msec for both.
> 
> I would like to request feedback on this approach; I think this is
> better than the "hardcoded steps" as before, but maybe someone can come
> up with an ever better idea....

logarithms pop to mind (again :-)

> static unsigned long __estimate_accuracy(struct timespec *tv)
> {
>         unsigned long slack;
>         int divfactor = 1000;
>   
>         if (task_nice(current))

This triggers for both -nice and +nice tasks, it might be worth
differentiating between those.

>                 divfactor = divfactor / 5;
> 
>         slack = tv->tv_nsec / divfactor; 
>         slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
> 
>         if (slack > 100 * NSEC_PER_MSEC)
>                 slack =  100 * NSEC_PER_MSEC;
>         return slack;
> }
> 
> static unsigned long estimate_accuracy(struct timespec *tv)
> {
>         unsigned long ret;
>         struct timespec now;
> 
>         /*
>          * Realtime tasks get a slack of 0 for obvious reasons.
>          */
> 
>         if (current->policy == SCHED_FIFO ||
>                 current->policy == SCHED_RR)
>                 return 0;

rt_task(current) ?

>         ktime_get_ts(&now);
>         now = timespec_sub(*tv, now);
>         ret = __estimate_accuracy(&now);
>         if (ret < current->timer_slack_ns)
>                 return current->timer_slack_ns;

pull out the max from __estimate_accuracy() and use clamp() here?

>         return ret;
> }

Re: update on hrtimer based select/poll and range-hrtimers

[Arjan van de Ven]

On Sun, 07 Sep 2008 22:46:39 +0200
Peter Zijlstra <peterz@infradead.org> wrote:
> > "0.1% of the time" for not-nice, not realtime tasks
> > "0.5% of the time" for nice, not realtime tasks
> > with a cap of 100msec for both.
> > 
> > I would like to request feedback on this approach; I think this is
> > better than the "hardcoded steps" as before, but maybe someone can
> > come up with an ever better idea....
> 
> logarithms pop to mind (again :-)

and it still doesn't strike me as "oh yeah"

how would you see this work? Can you give a numeric example?

> 
> > static unsigned long __estimate_accuracy(struct timespec *tv)
> > {
> >         unsigned long slack;
> >         int divfactor = 1000;
> >   
> >         if (task_nice(current))
> 
> This triggers for both -nice and +nice tasks, it might be worth
> differentiating between those.

good spotting; will fix.


> > 
> >         if (current->policy == SCHED_FIFO ||
> >                 current->policy == SCHED_RR)
> >                 return 0;
> 
> rt_task(current) ?

hmm that's not exactly equivalent.. I can see either but still.


-- 
If you want to reach me at my work email, use arjan@linux.intel.com
For development, discussion and tips for power savings, 
visit http://www.lesswatts.org

Re: update on hrtimer based select/poll and range-hrtimers

[Ingo Molnar]

* Arjan van de Ven <arjan@infradead.org> wrote:

> Hi,
> 
> since the last lkml posting I've merged a few fixes and added comments
> from Peter, and I've redone the "estimate_accuracy" function.
> 
> Rather than reposting the entire series, I'll point to the git tree at
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/arjan/linux-2.6-hrtimer.git

pulled into tip/timers/range-hrtimers, thanks Arjan!

	Ingo