[Devel] Re: Cgroup OOM bug

KAMEZAWA Hiroyuki kamezawa.hiroyu at jp.fujitsu.com
Sun May 9 17:01:31 PDT 2010


On Fri, 7 May 2010 14:30:21 -0700
Jeremy Hanmer <jeremy at hq.newdream.net> wrote:

> I've found a reproducible case where you can create an infinite loop inside the oom killer if you create a cgroup with a single process that has oom_adj=-17 and trigger an OOM situation inside that cgroup.
> 
> It seems that oom_adj should probably be ignored in the case where it makes the OOM situation unsolvable.  Here's an excerpt from the console output when the bug is triggered:
> 
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> Memory cgroup out of memory: kill process 25306 (memeater) score 0 or a child
> 

Now, In -mm tree, 2 options are supproted and it's under test.

  1. oom_notifier ... an event notifier which notifies OOM.

  2. oom_disable  ... disable OOM-Killer. All threads will wait for the end of
     OOM. No cpu consumption.

By 1+2, an user(user-land deamon) can implement its own OOM handler.
Please wait it.

Thanks,
-Kame

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers




More information about the Devel mailing list