linux-kernel - Re: [PATCH v2 5/6] sched/rt: Better manage pushing unfit tasks on wakeup

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: <20200225035505.GI28029@codeaurora.org>
Date:   Tue, 25 Feb 2020 09:25:05 +0530
From:   Pavan Kondeti <pkondeti@...eaurora.org>
To:     Qais Yousef <qais.yousef@....com>
Cc:     Ingo Molnar <mingo@...nel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Juri Lelli <juri.lelli@...hat.com>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2 5/6] sched/rt: Better manage pushing unfit tasks on
 wakeup

On Mon, Feb 24, 2020 at 05:41:39PM +0000, Qais Yousef wrote:
> On 02/24/20 21:34, Pavan Kondeti wrote:
> > Hi Qais,
> > 
> > On Mon, Feb 24, 2020 at 5:42 PM Qais Yousef <qais.yousef@....com> wrote:
> > [...]
> > > We could do, temporarily, to get these fixes into 5.6. But I do think
> > > select_task_rq_rt() doesn't do a good enough job into pushing unfit tasks to
> > > the right CPUs.
> > >
> > > I don't understand the reasons behind your objection. It seems you think that
> > > select_task_rq_rt() should be enough, but not AFAICS. Can you be a bit more
> > > detailed please?
> > >
> > > FWIW, here's a screenshot of what I see
> > >
> > >         https://imgur.com/a/peV27nE
> > >
> > > After the first activation, select_task_rq_rt() fails to find the right CPU
> > > (due to the same move all tasks to the cpumask_fist()) - but when the task
> > > wakes up on 4, the logic I put causes it to migrate to CPU2, which is the 2nd
> > > big core. CPU1 and CPU2 are the big cores on Juno.
> > >
> > > Now maybe we should fix select_task_rq_rt() to better balance tasks, but not
> > > sure how easy is that.
> > >
> > 
> > Thanks for the trace. Now things are clear to me. Two RT tasks woke up
> > simultaneously and the first task got its previous CPU i.e CPU#1. The next task
> > goes through find_lowest_rq() and got the same CPU#1. Since this task priority
> > is not more than the just queued task (already queued on CPU#1), it is sent
> > to its previous CPU i.e CPU#4 in your case.
> > 
> > From task_woken_rt() path, CPU#4 attempts push_rt_tasks(). CPU#4 is
> > not overloaded,
> > but we have rt_task_fits_capacity() check which forces the push. Since the CPU
> > is not overloaded, your has_unfit_tasks() comes to rescue and push the
> > task. Since
> > the task has not scheduled in yet, it is eligible for push. You added checks
> > to skip resched_curr() in push_rt_tasks() otherwise the push won't happen.
> 
> Nice summary, that's exactly what it is :)
> 
> > Finally, I understood your patch. Obviously this is not clear to me
> > before. I am not
> > sure if this patch is the right approach to solve this race. I will
> > think a bit more.
> 
> I haven't been staring at this code for as long as you, but since we have
> logic at wakeup to do a push, I think we need something here anyway for unfit
> tasks.
> 
> Fixing select_task_rq_rt() to better balance tasks will help a lot in general,
> but if that was enough already then why do we need to consider a push at the
> wakeup at all then?
> 
> AFAIU, in SMP the whole push-pull mechanism is racy and we introduce redundancy
> at taking the decision on various points to ensure we minimize this racy nature
> of SMP systems. Anything could have happened between the time we called
> select_task_rq_rt() and the wakeup, so we double check again before we finally
> go and run. That's how I interpret it.
> 
> I am open to hear about other alternatives first anyway. Your help has been
> much appreciated so far.
> 

The search inside find_lowest_rq() happens without any locks so I believe it
is expected to have races like this. In fact there is a comment in the code
saying "This test is optimistic, if we get it wrong the load-balancer
will have to sort it out" in select_task_rq_rt(). However, the push logic
as of today works only for overloaded case. In that sense, your patch fixes
this race for b.L systems. At the same time, I feel like tracking nonfit tasks
just to fix this race seems to be too much. I will leave this to Steve and
others to take a decision.

I thought of suggesting to remove the below check from select_task_rq_rt()

p->prio < cpu_rq(target)->rt.highest_prio.curr

which would then make the target CPU overloaded and the push logic would
spread the tasks. That works for a b.L system too. However there seems to
be a very good reason for doing this. see
https://lore.kernel.org/patchwork/patch/539137/

The fact that a CPU is part of lowest_mask but running a higher prio RT
task means there is a race. Should we retry one more time to see if we find
another CPU?

Thanks,
Pavan

-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.