Re: [patch 12/15] sched: maintain throttled rqs as a list

[Paul Turner <pjt@google.com>]

On Thu, Apr 21, 2011 at 7:50 PM, Hidetoshi Seto
<seto.hidetoshi@jp.fujitsu.com> wrote:
> Hi Paul,
>
> (2011/03/23 12:03), Paul Turner wrote:
>> @@ -1434,7 +1435,11 @@ static void throttle_cfs_rq(struct cfs_r
>>       if (!se)
>>               rq->nr_running += task_delta;
>>
>> +     raw_spin_lock(&cfs_b->lock);
>>       cfs_rq->throttled = 1;
>> +     list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
>> +     raw_spin_unlock(&cfs_b->lock);
>> +
>>       cfs_rq->throttled_timestamp = rq_of(cfs_rq)->clock;
>>  }
>>
>
> Though I'm looking forward to see your updates soon, just FYI,
> I found that somehow throttle_cfs_rq() was called with already
> throttled cfs_rq.
>
> So it breaks list throttled_cfs_rq by repeating list_add_tail_rcu()
> and results in lockups like that Bharata and Xiao already reported.
>
> There should be better fix, but following first aid to put_prev_entity()
> worked for me:
>
> -       if (!within_bandwidth(cfs_rq))
> +       if (!within_bandwidth(cfs_rq) && !cfs_rq->throttled)
>                throttle_cfs_rq(cfs_rq);
>        else
>                return_cfs_rq_quota(cfs_rq);
>
> I believe you can do better next time.
>

Aha!

Good catch -- Thanks Hidetoshi!  I've been looking for that one.

I'm figuring out exactly how this came about and I will repost.

Thanks!

>
> Thanks,
> H.Seto
>
>