linux-kernel - Re: [PATCH 7/8] sched/fair: Attempt misfit active balance when migration_type != migrate

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: <CAKfTPtBSkg2tBH5U+GvRWHBe3qxCqdB3w6WPJsZPADDbVQCc0w@mail.gmail.com>
Date:   Tue, 9 Feb 2021 09:58:59 +0100
From:   Vincent Guittot <vincent.guittot@...aro.org>
To:     Valentin Schneider <valentin.schneider@....com>
Cc:     linux-kernel <linux-kernel@...r.kernel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...nel.org>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Morten Rasmussen <morten.rasmussen@....com>,
        Qais Yousef <qais.yousef@....com>,
        Quentin Perret <qperret@...gle.com>,
        Pavan Kondeti <pkondeti@...eaurora.org>,
        Rik van Riel <riel@...riel.com>
Subject: Re: [PATCH 7/8] sched/fair: Attempt misfit active balance when
 migration_type != migrate_misfit

On Thu, 28 Jan 2021 at 19:32, Valentin Schneider
<valentin.schneider@....com> wrote:
>
> Consider a 4-CPU big.LITTLE system with CPUs 0-1 as LITTLEs and CPUs 2-3 as
> bigs. The resulting sched_domain hierarchy is:
>
>   DIE [          ]
>   MC  [    ][    ]
>        0  1  2  3
>
> When running a multithreaded CPU-bound workload (i.e. 1 hog per CPU), the
> expected behaviour is to have the about-to-idle big CPUs pull a hog from
> the LITTLEs, since bigs will complete their work sooner than LITTLEs.
>
> Further Consider a scenario where:
> - CPU0 is idle (e.g. its hog got migrated to one of the big CPUs)
> - CPU1 is currently executing a per-CPU kworker, preempting the CPU hog
> - CPU2 and CPU3 are executing CPU-hogs
>
> CPU0 goes through load_balance() at MC level, and tries to pick stuff from
> CPU1, but:
> - the hog can't be pulled, because it's task_hot()
> - the kworker can't be pulled, because it's pinned to CPU1, which sets
>   LBF_SOME_PINNED
>
> This load balance attempts ends with no load pulled, LBF_SOME_PINNED set,
> and as a consequence we set the imbalance flag of DIE's [0, 1]
> sched_group_capacity.
>
> Shortly after, CPU2 completes its work and is about to go idle. It goes
> through the newidle_balance(), and we would really like it to active
> balance the hog running on CPU1 (which is a misfit task). However,
> sgc->imbalance is set for the LITTLE group at DIE level, so the group gets
> classified as group_imbalanced rather than group_misfit_task.
>
> Unlike group_misfit_task (via migrate_misfit), the active balance logic
> doesn't have any specific case for group_imbalanced, so CPU2 ends up going
> idle. We'll have to wait for a load balance on CPU0 or CPU1 to happen and
> clear the imbalance flag, and then for another DIE-level load-balance on
> CPU2 to happen to pull the task off of CPU1. That's several precious
> milliseconds wasted down the drain.
>
> Giving group_misfit_task a higher group_classify() priority than
> group_imbalance doesn't seem like the right thing to do. Instead, make

Could you explain why ?

> need_active_balance() return true for any migration_type when the
> destination CPU is idle and the source CPU has a misfit task.
>
> While at it, add an sd_has_asym_cpucapacity() guard in
> need_active_balance().
>
> Signed-off-by: Valentin Schneider <valentin.schneider@....com>
> ---
>  kernel/sched/fair.c | 13 +++++++++++++
>  1 file changed, 13 insertions(+)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 0ac2f876b86f..cba9f97d9beb 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -9557,9 +9557,22 @@ static int need_active_balance(struct lb_env *env)
>                         return 1;
>         }
>
> +       if (!sd_has_asym_cpucapacity(sd))
> +               return 0;
> +
>         if (env->migration_type == migrate_misfit)
>                 return 1;
>
> +       /*
> +        * If we failed to pull anything and the src_rq has a misfit task, but
> +        * the busiest group_type was higher than group_misfit_task, try to
> +        * go for a misfit active balance anyway.
> +        */
> +       if ((env->idle != CPU_NOT_IDLE) &&
> +           env->src_rq->misfit_task_load &&
> +           cpu_capacity_greater(env->dst_cpu, env->src_cpu))
> +               return 1;
> +
>         return 0;
>  }
>
> --
> 2.27.0
>