Wednesday, August 5, 2009

Move oom_adj to signal_struct discussion

I copy and paste the content here for future reference:

Full discussion thread is here:
http://marc.info/?l=linux-mm&m=124938152105260&w=2


http://marc.info/?l=linux-mm&m=124938156605311&w=2



Minchan Kim wrote:
> 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 ?
>

Current code is as following
==
do_each_thread(g,p) {
......
p = badness();

record p of highest badness.
}
p = higest badness thread.

Scan all threads which shares mm_struct of p. and check oom_adj

==
Assume a process which has 20000 threads. And 1 of thread has OOM_DISABLE.

Then, at worst, this scan will needs
(1+2+3+....+20000) * (20000-1) scan. (when ignoring other processes)
even with your patch.

This means the kernel wastes enough long time that Cluster-Management-Software can
detetct this as livelock, and do reboot/cluster-fail-over.

Fixing livelock is not the last goal. I (we) would like to reduct stall time
to reasonable level. If we move oom_adj to signal_struct or mm_struct, scan-cost
will be only 20000. No retry at all.

And, if we can use for_each_process() rather than do_each_thread(),
scan-cost will be 1.

(BTW, "signal" struct is bad name I think, it should be "process" struct ;)


Thanks,
-Kame



> > > > > > > > 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.

No comments:

Post a Comment

Followers