Re: [PATCH 1/4] oom: move oom_adj to signal_struct

[Minchan Kim]

On Wed,  5 Aug 2009 15:53:31 +0900 (JST)
KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:

> > On Wed, 5 Aug 2009 15:03:23 +0900
> > KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > 
> > > On Wed, 5 Aug 2009 14:55:16 +0900
> > > Minchan Kim <minchan.kim@xxxxxxxxx> wrote:
> > > 
> > > > On Wed,  5 Aug 2009 11:51:31 +0900 (JST)
> > > > KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
> > > > 
> > > > > > On Wed,  5 Aug 2009 11:29:34 +0900 (JST)
> > > > > > KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
> > > > > > 
> > > > > > > Hi
> > > > > > > 
> > > > > > > > Hi, Kosaki. 
> > > > > > > > 
> > > > > > > > I am so late to invole this thread. 
> > > > > > > > But let me have a question. 
> > > > > > > > 
> > > > > > > > What's advantage of placing oom_adj in singal rather than task ?
> > > > > > > > I mean task->oom_adj and task->signal->oom_adj ?
> > > > > > > > 
> > > > > > > > I am sorry if you already discussed it at last threads. 
> > > > > > > 
> > > > > > > Not sorry. that's very good question.
> > > > > > > 
> > > > > > > I'm trying to explain the detailed intention of commit 2ff05b2b4eac
> > > > > > > (move oom_adj to mm_struct).
> > > > > > > 
> > > > > > > In 2.6.30, OOM logic callflow is here.
> > > > > > > 
> > > > > > > __out_of_memory
> > > > > > > 	select_bad_process		for each task
> > > > > > > 		badness			calculate badness of one task
> > > > > > > 	oom_kill_process		search child
> > > > > > > 		oom_kill_task		kill target task and mm shared tasks with it
> > > > > > > 
> > > > > > > example, process-A have two thread, thread-A and thread-B and it 
> > > > > > > have very fat memory.
> > > > > > > And, each thread have following likes oom property.
> > > > > > > 
> > > > > > > 	thread-A: oom_adj = OOM_DISABLE, oom_score = 0
> > > > > > > 	thread-B: oom_adj = 0,           oom_score = very-high
> > > > > > > 
> > > > > > > Then, select_bad_process() select thread-B, but oom_kill_task refuse
> > > > > > > kill the task because thread-A have OOM_DISABLE.
> > > > > > > __out_of_memory() call select_bad_process() again. but select_bad_process()
> > > > > > > select the same task. It mean kernel fall in the livelock.
> > > > > > > 
> > > > > > > The fact is, select_bad_process() must select killable task. otherwise
> > > > > > > OOM logic go into livelock.
> > > > > > > 
> > > > > > > Is this enough explanation? thanks.
> > > > > > > 
> > > > 
> > > > The problem resulted from David patch.
> > > > It can solve live lock problem but make a new problem like vfork problem. 
> > > > I think both can be solved by different approach. 
> > > > 
> > > > It's just RFC. 
> > > > 
> > > > If some process is selected by OOM killer but it have a child of OOM immune,
> > > > We just decrease point of process. It can affect selection of bad process. 
> > > > After some trial, at last bad score is drastically low and another process is 
> > > > selected by OOM killer. So I think Live lock don't happen. 
> > > > 
> > > > New variable adding in task struct is rather high cost. 
> > > > But i think we can union it with oomkilladj 
> > > > since oomkilladj is used to present just -17 ~ 15. 
> > > > 
> > > > What do you think about this approach ?
> > > > 
> > > keeping this in "task" struct is troublesome.
> > > It may not livelock but near-to-livelock state, in bad case.
> > 
> > Hmm. I can't understand why it is troublesome. 
> > I think it's related to moving oom_adj to singal_struct. 
> > Unfortunately, I can't understand why we have to put oom_adj 
> > in singal_struct?
> > 
> > That's why I have a question to Kosaki a while ago. 
> > I can't understand it still. :-(
> > 
> > Could you elaborate it ?
> 
> Maybe, It's because my explanation is still poor. sorry.
> Please give me one more chance.
> 
> In my previous mail, I explained select_bad_process() must not
> unkillable task, is this ok?
> IOW, if all thread have the same oom_adj, the issue gone.
> 
> signal_struct is shared all thread in the process. then, the issue gone.
> 

Your and Kame's good explanation opens my eyes. :)
I realized your approach's benefit. 

Yes. Let's wait to listen others's opinios.



-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/