linux-kernel - Re: [PATCH 2/3] sched/fair: Fix newidle

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAKfTPtCJPHfoiuspYMGPARHdOuLJ6g0oUx2EQjdEPz729NrDPA@mail.gmail.com>
Date:   Wed, 22 Dec 2021 09:14:44 +0100
From:   Vincent Guittot <vincent.guittot@...aro.org>
To:     Vincent Donnefort <Vincent.Donnefort@....com>
Cc:     peterz@...radead.org, mingo@...hat.com,
        linux-kernel@...r.kernel.org, dietmar.eggemann@....com,
        Valentin.Schneider@....com, Morten.Rasmussen@....com,
        qperret@...gle.com
Subject: Re: [PATCH 2/3] sched/fair: Fix newidle_balance() for overutilized systems

On Mon, 20 Dec 2021 at 12:43, Vincent Donnefort
<vincent.donnefort@....com> wrote:
>
> On Energy-Aware Scheduling systems, load balancing is disabled in favor of
> energy based placement, until one of the CPU is identified as being
> overutilized. Once the overutilization is resolved, two paths can lead to
> marking the system as non overutilized again:
>
>   * load_balance() triggered from newidle_balance().
>   * load_balance() triggered from the scheduler tick.
>
> However, small caveat for each of those paths. newidle_balance() needs
> rd->overload set to run load_balance(), while the load_balance() triggered
> by the scheduler tick needs to run from the first idle CPU of the root
> domain (see should_we_balance()).
>
> Overutilized can be triggered without setting overload (this can happen
> for a CPU which had a misfit task but didn't had its util_avg updated
> yet). Then, only the scheduler tick could help to reset overutilized...
> but if most of the CPUs are idle, it is very unlikely load_balance() would
> run on the only CPU which can reset the flag. This means the root domain

AFAICT, this will happen as you don't have to run on the only CPU but
for the only CPU and this is what ilb is doing. So your problem is not
to run on the only CPU that can clear overutilized

> can spuriously maintain overutilized for a long period of time.
>
> We then need newidle_balance() to proceed with balancing if the system is
> overutilized.

Always triggering a costly newidle_balance when you are already
overutilized for the sole purpose of clearing overutilized seems to be
a bit overkill.

Furthermore, nothing prevents us to abort newidle_balance before
reaching the root domain

So this doesn't seem like the good way to proceed

>
> Fixes: 2802bf3cd936 ("sched/fair: Add over-utilization/tipping point indicator")
> Signed-off-by: Vincent Donnefort <vincent.donnefort@....com>
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e2f6fa14e5e7..51f6f55abb37 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10849,7 +10849,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
>         rcu_read_lock();
>         sd = rcu_dereference_check_sched_domain(this_rq->sd);
>
> -       if (!READ_ONCE(this_rq->rd->overload) ||
> +       if ((!READ_ONCE(this_rq->rd->overload) &&
> +           !READ_ONCE(this_rq->rd->overutilized)) ||
>             (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
>
>                 if (sd)
> --
> 2.25.1
>