Re: [PATCH 2/4] oom: kill younger process first

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, May 12, 2011 at 01:17:13PM +0900, Minchan Kim wrote:
> On Thu, May 12, 2011 at 12:39 PM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > On Thu, 12 May 2011 11:23:38 +0900
> > Minchan Kim <minchan.kim@gmail.com> wrote:
> >
> >> On Thu, May 12, 2011 at 10:53 AM, KAMEZAWA Hiroyuki
> >> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> >> > On Thu, 12 May 2011 10:30:45 +0900
> >> > Minchan Kim <minchan.kim@gmail.com> wrote:
> >
> >> > As above implies, (B)->prev pointer is invalid pointer after list_del().
> >> > So, there will be race with list modification and for_each_list_reverse under
> >> > rcu_read__lock()
> >> >
> >> > So, when you need to take atomic lock (as tasklist lock is) is...
> >> >
> >> >  1) You can't check 'entry' is valid or not...
> >> >    In above for_each_list_rcu(), you may visit an object which is under removing.
> >> >    You need some flag or check to see the object is valid or not.
> >> >
> >> >  2) you want to use list_for_each_safe().
> >> >    You can't do list_del() an object which is under removing...
> >> >
> >> >  3) You want to walk the list in reverse.
> >> >
> >> >  3) Some other reasons. For example, you'll access an object pointed by the
> >> >    'entry' and the object is not rcu safe.
> >> >
> >> > make sense ?
> >>
> >> Yes. Thanks, Kame.
> >> It seems It is caused by prev poisoning of list_del_rcu.
> >> If we remove it, isn't it possible to traverse reverse without atomic lock?
> >>
> >
> > IIUC, it's possible (Fix me if I'm wrong) but I don't like that because of 2 reasons.
> >
> > 1. LIST_POISON is very important information at debug.
> 
> Indeed.
> But if we can get a better something although we lost debug facility,
> I think it would be okay.
> 
> >
> > 2. If we don't clear prev pointer, ok, we'll allow 2 directional walk of list
> >   under RCU.
> >   But, in following case
> >   1. you are now at (C). you'll visit (C)->next...(D)
> >   2. you are now at (D). you want to go back to (C) via (D)->prev.
> >   3. But (D)->prev points to (B)
> >
> >  It's not a 2 directional list, something other or broken one.
> 
> Yes. but it shouldn't be a problem in RCU semantics.
> If you need such consistency, you should use lock.
> 
> I recall old thread about it.
> In http://lwn.net/Articles/262464/, mmutz and Paul already discussed
> about it. :)
> 
> >  Then, the rculist is 1 directional list in nature, I think.
> 
> Yes. But Why RCU become 1 directional list is we can't find a useful usecases.
> 
> >
> > So, without very very big reason, we should keep POISON.
> 
> Agree.
> I don't insist on it as it's not a useful usecase for persuading Paul.
> That's because it's not a hot path.
> 
> It's started from just out of curiosity.
> Thanks for very much clarifying that, Kame!

Indeed, we would need a large performance/scalability/simplicity advantage
to put up with such a loss of debugging information.  If it turns out
that you really need this, please let me know, but please also provide
data supporting your need.

							Thanx, Paul