Received-SPF: pass (google.com: domain of linux-kernel+bounces-12773-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Thu, 28 Dec 2023 16:21:32 +0000
From: Qais Yousef <qyousef@layalina.io>
To: John Stultz <jstultz@google.com>
Cc: LKML <linux-kernel@vger.kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Joel Fernandes <joelaf@google.com>,
	Qais Yousef <qyousef@google.com>, Ingo Molnar <mingo@redhat.com>,
	Juri Lelli <juri.lelli@redhat.com>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Valentin Schneider <vschneid@redhat.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>,
	Zimuzo Ezeozue <zezeozue@google.com>,
	Youssef Esmat <youssefesmat@google.com>,
	Mel Gorman <mgorman@suse.de>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	Will Deacon <will@kernel.org>, Waiman Long <longman@redhat.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	"Paul E . McKenney" <paulmck@kernel.org>, kernel-team@android.com,
	Mike Galbraith <efault@gmx.de>,
	Daniel Bristot de Oliveira <bristot@kernel.org>
Subject: Re: [PATCH v6 01/20] sched: Unify runtime accounting across classes
Message-ID: <20231228162132.mauucgalrcatvwxn@airbuntu>
References: <20231106193524.866104-1-jstultz@google.com>
 <20231106193524.866104-2-jstultz@google.com>
 <20231217161927.vewygmvxvddhxs65@airbuntu>
 <CANDhNCraerdXXpxZ4nDB=aaNmmMnH+2M2yQ=NkFAWECOX6T5CA@mail.gmail.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CANDhNCraerdXXpxZ4nDB=aaNmmMnH+2M2yQ=NkFAWECOX6T5CA@mail.gmail.com>

On 12/18/23 12:23, John Stultz wrote:
> On Sun, Dec 17, 2023 at 8:19 AM Qais Yousef <qyousef@layalina.io> wrote:
> > On 11/06/23 19:34, John Stultz wrote:
> > > From: Peter Zijlstra <peterz@infradead.org>
> > >
> > > All classes use sched_entity::exec_start to track runtime and have
> > > copies of the exact same code around to compute runtime.
> > >
> > > Collapse all that.
> > >
> ...
> > Looks like this actually got merged into tip via the deadline server work :-)
> 
> Oh! That's great to see! The patch has been floating around for a while.
> 
> > Though not sure if I caught a bug here
> >
> > > diff --git a/kernel/sched/stop_task.c b/kernel/sched/stop_task.c
> > > index 85590599b4d6..7595494ceb6d 100644
> > > --- a/kernel/sched/stop_task.c
> > > +++ b/kernel/sched/stop_task.c
> > > @@ -70,18 +70,7 @@ static void yield_task_stop(struct rq *rq)
> > >
> > >  static void put_prev_task_stop(struct rq *rq, struct task_struct *prev)
> > >  {
> > > -     struct task_struct *curr = rq->curr;
> > > -     u64 now, delta_exec;
> > > -
> > > -     now = rq_clock_task(rq);
> > > -     delta_exec = now - curr->se.exec_start;
> > > -     if (unlikely((s64)delta_exec < 0))
> > > -             delta_exec = 0;
> >
> > If negative instead of returning for stopper task; we set delta_exec to 0
> >
> > > -
> > > -     schedstat_set(curr->stats.exec_max,
> > > -                   max(curr->stats.exec_max, delta_exec));
> > > -
> > > -     update_current_exec_runtime(curr, now, delta_exec);
> >
> > And curry on to do time accounting
> >
> > > +     update_curr_common(rq);
> >
> > But the new function will return early without doing accounting. Wouldn't this
> > re-introrduce 8f6189684eb4 ("sched: Fix migration thread runtime bogosity")?
> 
> Hrm. So first, good eye for catching this!
> Looking through the code, much of the accounting logic we end up
> skipping doesn't have much effect when delta_exec = 0, so it seems
> mostly harmless to return early without the accounting.
> 
> Though, there is one side-effect that does get skipped, which is the
> removed update_current_exec_runtime() unconditionally sets:
>   curr->se.exec_start = now;
> 
> Which basically resets the accounting window.
> 
> From the commit, It's unclear how intentional this side-effect is for
> the edge case where the interval is negative.
> 
> I can't say I've really wrapped my head around the cases where the
> se.exec_start would get ahead of the rq_clock_task(), so it's not
> clear in which cases we would want to reset the accounting window vs
> wait for the rq_clock_task() to catch up.  But as this is getting
> called from put_prev_task_stop(), it seems we're closing the
> accounting window here anyway, and later set_next_task_stop() would be
> called (which sets se.exec_start, resetting the accounting) to start
> the accounting window again.
> 
> So you are right that there is a practical change in behavior, but I
> don't think I see it having an effect.

Yes, agreed. I couldn't reproduce any problem and I can't see a terrible side
effect of returning early as well compared to continuing to do the accounting.


Cheers

--
Qais Yousef