Re: [patch -mm v2] mm: introduce oom_adj_child

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

One mistake.

> > > And, May I explay why I think your oom_adj_child is wrong idea?
> > > The fact is: new feature introducing never fix regression. yes, some
> > > application use new interface and disappear the problem. but other
> > > application still hit the problem. that's not correct development style
> > > in kernel.
> > > 
> > 
> > So you're proposing that we forever allow /proc/pid/oom_score to be 
> > completely wrong for pid without any knowledge to userspace?  That we 
> > falsely advertise what it represents and allow userspace to believe that 
> > changing oom_adj for a thread sharing memory with other threads actually 
> > changes how the oom killer selects tasks?
> 
> No. perhaps no doublly.
> 
> 1) In my patch, oom_score is also per-process value. all thread have the same
>    oom_score.
>    It's clear meaning.

it's wrong explanation. oom_score is calculated from the same oom_adj.
but it have each different oom_score. sorry my confused.



> 2) In almost case, oom_score display collect value because oom_adj is per-process
>    value too. 
>    Yes, there is one exception. vfork() and change oom_adj'ed process might display 
>    wrong value. but I don't think it is serious problem because vfork() process call
>    exec() soon.
>    Administrator never recognize this difference.
> 
> > Please.
> 
> David, I hope you join to fix this regression. I can't believe we
> can't fix this issue honestly.
> 
> 
>