Received-SPF: pass (google.com: domain of linux-kernel+bounces-215699-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Precedence: bulk
MIME-Version: 1.0
References: <20240613015837.4132703-1-jstultz@google.com> <20240613100441.GC17707@noisy.programming.kicks-ass.net>
 <20240613115142.kxrmlf3btmwjcprg@airbuntu> <20240614094833.GM8774@noisy.programming.kicks-ass.net>
In-Reply-To: <20240614094833.GM8774@noisy.programming.kicks-ass.net>
From: John Stultz <jstultz@google.com>
Date: Fri, 14 Jun 2024 21:30:13 -0700
Message-ID: <CANDhNCrVVTtp37BH1zdTBUJqXCFpzSAOpqEvXii2kKTg+k=wBQ@mail.gmail.com>
Subject: Re: [PATCH] RFC: sched: Rework task_sched_runtime to avoid calling update_rq_clock
To: Peter Zijlstra <peterz@infradead.org>
Cc: Qais Yousef <qyousef@layalina.io>, LKML <linux-kernel@vger.kernel.org>, 
	Thomas Gleixner <tglx@linutronix.de>, Frederic Weisbecker <frederic@kernel.org>, Ingo Molnar <mingo@redhat.com>, 
	Juri Lelli <juri.lelli@redhat.com>, Vincent Guittot <vincent.guittot@linaro.org>, 
	Dietmar Eggemann <dietmar.eggemann@arm.com>, Steven Rostedt <rostedt@goodmis.org>, 
	Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>, 
	Daniel Bristot de Oliveira <bristot@redhat.com>, Valentin Schneider <vschneid@redhat.com>, 
	Joel Fernandes <joel@joelfernandes.org>, kernel-team@android.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Fri, Jun 14, 2024 at 2:48=E2=80=AFAM Peter Zijlstra <peterz@infradead.or=
g> wrote:
> On Thu, Jun 13, 2024 at 12:51:42PM +0100, Qais Yousef wrote:
> > On 06/13/24 12:04, Peter Zijlstra wrote:
> > > @@ -5459,6 +5458,8 @@ void sched_tick(void)
> > >
> > >     sched_clock_tick();
> > >
> > > +   psi_account_irqtime(curr, &rq->psi_irq_time);
> > > +
> >
> > If wakeup preemption causes a context switch, wouldn't we lose this
> > information then? I *think* active migration might cause this informati=
on to be
> > lost too.
>
> I'm not sure what would be lost ?! the accounting is per cpu, not per
> task afaict. That said,...
>
> > pick_next_task() might be a better place to do the accounting?
>
> Additionally, when there has been an effective cgroup switch. Only on
> switch doesn't work for long running tasks, then the PSI information
> will be artitrarily long out of date.
>
> Which then gets me something like the (completely untested) below..
>
> Hmm?

I applied and booted with this. It still takes the accounting out of
the hotpath for the CLOCK_THREAD_CPUTIME_ID the microbenchmark
performance is back to 5.10 numbers.

I don't have any correctness tests for irqtime measurements, so I'll
have to try to work something up for that next week.

thanks
-john