lkml.org 
[lkml]   [2018]   [Jul]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] sched/fair: Avoid divide by zero when rebalancing domains
From
Date
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.

\
 
 \ /
  Last update: 2018-07-05 18:54    [W:0.082 / U:0.824 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site