Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH] sched/fair: Avoid divide by zero when rebalancing domains
To:     Matt Fleming <matt@codeblueprint.co.uk>
Cc:     Peter Zijlstra <peterz@infradead.org>,
        linux-kernel@vger.kernel.org, Ingo Molnar <mingo@kernel.org>,
        Mike Galbraith <umgwanakikbuti@gmail.com>
References: <20180704142455.16035-1-matt@codeblueprint.co.uk>
 <55afee27-4143-e08c-b254-0d68a05d5ee6@arm.com>
 <20180705132726.GB3864@codeblueprint.co.uk>
From:   Valentin Schneider <valentin.schneider@arm.com>
Message-ID: <94149109-a54c-fc5d-7b56-e786c8de5b94@arm.com>
Date:   Thu, 5 Jul 2018 17:54:02 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101
 Thunderbird/52.8.0
MIME-Version: 1.0
In-Reply-To: <20180705132726.GB3864@codeblueprint.co.uk>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 05/07/18 14:27, Matt Fleming wrote:
> On Thu, 05 Jul, at 11:10:42AM, Valentin Schneider wrote:
>> Hi,
>>
>> On 04/07/18 15:24, Matt Fleming wrote:
>>> It's possible that the CPU doing nohz idle balance hasn't had its own
>>> load updated for many seconds. This can lead to huge deltas between
>>> rq->avg_stamp and rq->clock when rebalancing, and has been seen to
>>> cause the following crash:
>>>
>>>  divide error: 0000 [#1] SMP
>>>  Call Trace:
>>>   [<ffffffff810bcba8>] update_sd_lb_stats+0xe8/0x560

My confusion comes from not seeing where that crash happens. Would you mind
sharing the associated line number? I can feel the "how did I not see this"
from there but it can't be helped :(

>>>   [<ffffffff810bd04d>] find_busiest_group+0x2d/0x4b0
>>>   [<ffffffff810bd640>] load_balance+0x170/0x950
>>>   [<ffffffff810be3ff>] rebalance_domains+0x13f/0x290
>>>   [<ffffffff810852bc>] __do_softirq+0xec/0x300
>>>   [<ffffffff8108578a>] irq_exit+0xfa/0x110
>>>   [<ffffffff816167d9>] reschedule_interrupt+0xc9/0xd0
>>>
>>
>> Do you have some sort of reproducer for that crash? If not I guess I can cook
>> something up with a quiet userspace & rt-app, though I've never seen that one
>> on arm64.
>  
> Unfortunately no, I don't have a reproduction case. Would love to have
> one to verify the patch though.
> 
>>> Make sure we update the rq clock and load before balancing.
>>>
>>> Cc: Ingo Molnar <mingo@kernel.org>
>>> Cc: Mike Galbraith <umgwanakikbuti@gmail.com>
>>> Cc: Peter Zijlstra <peterz@infradead.org>
>>> Signed-off-by: Matt Fleming <matt@codeblueprint.co.uk>
>>> ---
>>>  kernel/sched/fair.c | 10 ++++++++++
>>>  1 file changed, 10 insertions(+)
>>>
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index 2f0a0be4d344..2c81662c858a 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -9597,6 +9597,16 @@ static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
>>>  	 */
>>>  	smp_mb();
>>>  
>>> +	/*
>>> +	 * Ensure this_rq's clock and load are up-to-date before we
>>> +	 * rebalance since it's possible that they haven't been
>>> +	 * updated for multiple schedule periods, i.e. many seconds.
>>> +	 */
>>> +	raw_spin_lock_irq(&this_rq->lock);
>>> +	update_rq_clock(this_rq);
>>> +	cpu_load_update_idle(this_rq);
>>> +	raw_spin_unlock_irq(&this_rq->lock);
>>> +
>>
>> I'm failing to understand why the updates further down below are seemingly
>> not enough. After we've potentially done 
>>
>>     update_rq_clock(rq);
>>     cpu_load_update_idle(rq);
>>
>> for all nohz cpus != this_cpu, we still end up doing:
>>
>>     if (idle != CPU_NEWLY_IDLE) {
>> 	    update_blocked_averages(this_cpu);
>> 	    has_blocked_load |= this_rq->has_blocked_load;
>>     }
>>
>> which should properly update this_rq's clock and load before we attempt to do
>> any balancing on it.
>  
> But cpu_load_update_idle() and update_blocked_averages() are not the same
> thing.
> 

Right, we don't do any rq->cpu_load[] update for this_rq in the current nohz
code (i.e. by using update_blocked_averages()), which I think we do want to
do. I'm just miserably failing to find how not doing this leads to a div by 0.