Date: Fri, 11 Apr 2014 16:35:12 +0200
From: Frederic Weisbecker <fweisbec@gmail.com>
To: Viresh Kumar <viresh.kumar@linaro.org>
Cc: Thomas Gleixner <tglx@linutronix.de>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Lists linaro-kernel <linaro-kernel@lists.linaro.org>
Subject: Re: [Query]: tick-sched: can idle_active be false in
 tick_nohz_idle_exit()?
Message-ID: <20140411143509.GB3438@localhost.localdomain>
References: <CAKohpon3vJzL51V9aOBcXKJC4nr3K-0q77HsxFi+=ACUqNqhcA@mail.gmail.com>
 <20140410145621.GC27654@localhost.localdomain>
 <CAKohpo=auNFsivuM21pDxJ0e4sP-ScWc9EPvzhT6qud_UBSJqw@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAKohpo=auNFsivuM21pDxJ0e4sP-ScWc9EPvzhT6qud_UBSJqw@mail.gmail.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org

On Fri, Apr 11, 2014 at 03:24:11PM +0530, Viresh Kumar wrote:
> On 10 April 2014 20:26, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > When a dynticks idle CPU is woken up (typically with an IPI), tick_nohz_stop_idle()
> > is called on interrupt entry but, because this is a waking up IPI, tick_nohz_start_idle()
> > won't be called. The reason is that need_resched() prevents tick_nohz_irq_exit() to be
> > called in irq_exit().
> >
> > After all if we know that the CPU is going to exit the idle task, we don't need to account
> > any more idle time. We also don't need to retry to enter in dynticks idle mode since we
> > are going to restart the tick with tick_nohz_idle_exit().
> >
> > So in case of wake up IPIs, we may end up with !ts->idle_active in tick_nohz_idle_exit() :)
> >
> > I must confess this is not obvious.
> 
> I agree.. I didn't had a clue that this can happen. Good that I asked
> this question :)
> 
> > It confused me as well when I met that part. A small
> > comment in tick_nohz_idle_exit() would be welcome ;)
> 
> Looks hard. I tried for a small comment and this is the smallest I could get:
> 
> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index c2d868d..26cf5bb 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -925,6 +925,22 @@ void tick_nohz_idle_exit(void)
> 
>         ts->inidle = 0;
> 
> +       /*
> +        * Can idle_active be false here?
> +        * Ideally this would be the sequence of calls:
> +        * - tick_nohz_idle_enter(), i.e. idle_active = true;
> +        * - local_irq_disable()
> +        * - IDLE
> +        * - wake up due to IPI or other interrupt
> +        * - local_irq_enable()
> +        * - tick_nohz_irq_enter(), i.e. idle_active = false;
> +        * - tick_nohz_irq_exit(), i.e. idle_active = true; This is not called
> +        *   in case of IPI's as need_resched() will prevent that in
> +        *   tick_irq_exit(), as we don't need to account any more for idle time
> +        *   or try to enter dyntics mode (We are going to exit idle state).
> +        *
> +        * - tick_nohz_idle_exit()
> +        */
>         if (ts->idle_active || ts->tick_stopped)
>                 now = ktime_get();

I'm sure we can summarize a lot of uninteresting details there. Many of what
you describe can be guessed after a simple review on the code. Let rather
focus on the trickies.

How about something like:

          /*
            * ts->active can be 0 if a wake up IPI pulled us out of idle mode. When
            * that happens we know we're exiting the idle task. Pending idle sleep
            * time is flushed on irq entry then no more is accounted afterward.
            * The need_resched() check on irq_exit() prevents from accounting more.
           */

> 
> 
> I am preparing a cleanup patchset (separate from the timer/hrtimers 36 patch
> set) for kernel/time/ and will add this change to that.. I am waiting
> for the merge
> window to close and Thomas to comment on my timers/hrtimers patches first.
> Only then will I send another 40 :)

Thanks! Note you can post before the merge window closes. Reviews are possible
anytime :)

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