Question : memrlimit cgroup's task_move (2.6.26-rc5-mm3)

Balbir Singh balbir at linux.vnet.ibm.com
Thu Jun 19 05:30:24 PDT 2008


KAMEZAWA Hiroyuki wrote:
> On Thu, 19 Jun 2008 12:24:29 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu at jp.fujitsu.com> wrote:
> 
>> On Thu, 19 Jun 2008 08:43:43 +0530
>> Balbir Singh <balbir at linux.vnet.ibm.com> wrote:
>>
>>>> I think the charge of the new group goes to minus. right ?
>>>> (and old group's charge never goes down.)
>>>> I don't think this is "no problem".
>>>>
>>>> What kind of patch is necessary to fix this ?
>>>> task_attach() should be able to fail in future ?
>>>>
>>>> I'm sorry if I misunderstand something or this is already in TODO list.
>>>>
>>> It's already on the TODO list. Thanks for keeping me reminded about it.
>>>
>> Okay, I'm looking foward to see how can_attach and roll-back(if necessary)
>> is implemnted.
>> As you know, I'm interested in how to handle failure of task move.
>>
> One more thing...
> Now, charge is done at
> 
>  - vm is inserted (special case?)
>  - vm is expanded (mmap is called, stack growth...)
> 
> And uncharge is done at
>  - vm is removed (success of munmap)
>  - exit_mm is called (exit of process)
> 
> But it seems charging at may_expand_vm() is not good.
> The mmap can fail after may_expand_vm() because of various reason,
> but charge is already done at may_expand_vm()....and no roll-back.
> 
> == an easy example of leak in stack growth handling ==
> [root at iridium kamezawa]# cat /opt/cgroup/test/memrlimit.usage_in_bytes
> 71921664
> [root at iridium kamezawa]# ulimit -s 3
> [root at iridium kamezawa]# ls
> Killed
> [root at iridium kamezawa]# ls
> Killed
> [root at iridium kamezawa]# ls
> Killed
> [root at iridium kamezawa]# ls
> Killed
> [root at iridium kamezawa]# ls
> Killed
> [root at iridium kamezawa]# ulimit -s unlimited
> [root at iridium kamezawa]# cat /opt/cgroup/test/memrlimit.usage_in_bytes
> 72368128
> [root at iridium kamezawa]#

Aaah.. I see.. I had it in place earlier, but moved them to may_expand_vm() on
review suggestions. I can move it out or try to unroll when things fail. I'll
experiment a bit more. Is there any particular method you prefer?

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL


More information about the Containers mailing list