On Wed 01-06-16 19:41:09, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Wed 01-06-16 00:53:03, Oleg Nesterov wrote:
> > > On 05/31, Michal Hocko wrote:
> > > >
> > > > Oleg has pointed out that can simplify both oom_adj_write and
> > > > oom_score_adj_write even further and drop the sighand lock. The only
> > > > purpose of the lock was to protect p->signal from going away but this
> > > > will not happen since ea6d290ca34c ("signals: make task_struct->signal
> > > > immutable/refcountable").
> > >
> > > Sorry for confusion, I meant oom_adj_read() and oom_score_adj_read().
> > >
> > > As for oom_adj_write/oom_score_adj_write we can remove it too, but then
> > > we need to ensure (say, using cmpxchg) that unpriviliged user can not
> > > not decrease signal->oom_score_adj_min if its oom_score_adj_write()
> > > races with someone else (say, admin) which tries to increase the same
> > > oom_score_adj_min.
> >
> > I am introducing oom_adj_mutex in a later patch so I will move it here.
>
> Can't we reuse oom_lock like
>
> if (mutex_lock_killable(&oom_lock))
> return -EINTR;
>
> ? I think that updating oom_score_adj unlikely races with OOM killer
> invocation, and updating oom_score_adj should be a killable operation.
We could but what would be an advantage? Do we really need a full oom
exclusion?
--
Michal Hocko
SUSE Labs
