linux-kernel - Re: [PATCH RFC] sched/fair: bump sd->max_newidle_lb

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: <CAKfTPtBE0_77+J-A7vWRKsHCCmuX1jWTbPYWGVPg1MYq_rv8Og@mail.gmail.com>
Date: Thu, 26 Jun 2025 16:26:00 +0200
From: Vincent Guittot <vincent.guittot@...aro.org>
To: Chris Mason <clm@...a.com>
Cc: Peter Zijlstra <peterz@...radead.org>, Chris Mason <clm@...com>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH RFC] sched/fair: bump sd->max_newidle_lb_cost when newidle
 balance fails

On Thu, 26 Jun 2025 at 12:58, Chris Mason <clm@...a.com> wrote:
>
> On 6/26/25 3:00 AM, Peter Zijlstra wrote:
> > On Tue, Jun 24, 2025 at 01:48:08PM -0700, Chris Mason wrote:
>
> [ ... ]
>
> > For the non-RFC version, please split this into a code move and a code
> > change -- I had to stare waaay to long to spot the difference (if we
> > keep this code movement at all).
>
> Sure
>
> >
> >>  /*
> >>   * Check this_cpu to ensure it is balanced within domain. Attempt to move
> >>   * tasks if there is an imbalance.
> >> @@ -11782,12 +11808,14 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
> >>
> >>      group = sched_balance_find_src_group(&env);
> >>      if (!group) {
> >> +            update_newidle_cost(sd, sd->max_newidle_lb_cost + sd->max_newidle_lb_cost / 2);
> >>              schedstat_inc(sd->lb_nobusyg[idle]);
> >>              goto out_balanced;
> >>      }
> >>
> >>      busiest = sched_balance_find_src_rq(&env, group);
> >>      if (!busiest) {
> >> +            update_newidle_cost(sd, sd->max_newidle_lb_cost + sd->max_newidle_lb_cost / 2);
> >>              schedstat_inc(sd->lb_nobusyq[idle]);
> >>              goto out_balanced;
> >>      }
> >
> > So sched_balance_rq() is used for pretty much all load-balancing, not
> > just newidle.
> >
> > Either make this conditional like:
> >
> >       if (idle == CPU_NEWLY_IDLE)
> >               update_newidle_cost(...);
> >
> > or do it all the callsite, where we find !pulled_task (ie failure).
> >
> > Specifically, we already do update_newidle_cost() there, perhaps inflate
> > the cost there instead?
> >
> >       if (!pulled_tasks)
> >               domain_cost += sysctl_sched_migration_cost;
>
> Got it, I'll play with that.  Vincent, was there a benchmark I can use
> to see if I've regressed the case you were focused on?

It's not a public benchmark but I had some unitary tests with tasks
waiting on a busy CPU while other CPUs become idle for a "long" time
(but still less than 500us in average). This is even more true with
frequency scaling which will minimize the idle duration by decreasing
the frequency

>
> -chris
>