Re: sched: circular dependency between sched_domains_mutex and oom_notify_list

[Sasha Levin <sasha.levin@oracle.com>]

On 02/18/2013 03:26 AM, Ingo Molnar wrote:
> 
> * Sasha Levin <sasha.levin@oracle.com> wrote:
> 
>> I suspect it's the result of adding the new rcu_oom_notify, but that happened
>> about half a year ago so I'm not sure why this showed up only now.
>>
>> [ 1039.634183] ======================================================
>> [ 1039.635717] [ INFO: possible circular locking dependency detected ]
>> [ 1039.637255] 3.8.0-rc7-next-20130215-sasha-00003-gea816fa #286 Tainted: G        W
>> [ 1039.639104] -------------------------------------------------------
>> [ 1039.640579] init/1 is trying to acquire lock:
>> [ 1039.641224]  ((oom_notify_list).rwsem){.+.+..}, at: [<ffffffff81141c8f>] __blocking_notifier_call_chain+0x7f/0xc0
> 
> We changed (optimized) rwsems via:
> 
>  3a15e0e0cdda rwsem: Implement writer lock-stealing for better scalability
> 
> so maybe it can hit different codepaths and races now?

I don't think that that patch would modify codepaths - it might just cause some
race conditions more probable.

I'm still thinking that the issue is the new rcu oom thingie and since this
codepath is very unlikely (running out of memory while sched_domains_mutex is
held - which is very rare) it's the culprit.


Thanks,
Sasha