On Fri, Aug 21, 2020 at 4:16 AM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>
> On 08/20, Eric W. Biederman wrote:
> >
> > That said if we are going for a small change why not:
> >
> > /*
> > * Make sure we will check other processes sharing the mm if this is
> > * not vfrok which wants its own oom_score_adj.
> > * pin the mm so it doesn't go away and get reused after task_unlock
> > */
> > if (!task->vfork_done) {
> > struct task_struct *p = find_lock_task_mm(task);
> >
> > if (p) {
> > - if (atomic_read(&p->mm->mm_users) > 1) {
> > + if (atomic_read(&p->mm->mm_users) > p->signal->nr_threads) {
>
> In theory this needs a barrier to avoid the race with do_exit(). And I'd
> suggest to use signal->live, I think signal->nr_threads should die...
> Something like
>
> bool probably_has_other_mm_users(tsk)
> {
> return atomic_read_acquire(&tsk->mm->mm_users) >
> atomic_read(&tsk->signal->live);
> }
>
> The barrier implied by _acquire ensures that if we race with the exiting
> task and see the result of exit_mm()->mmput(mm), then we must also see
> the result of atomic_dec_and_test(signal->live).
>
> Either way, if we want to fix the race with clone(CLONE_VM) we need other
> changes.
The way I understand this condition in __set_oom_adj() sync logic is
that we would be ok with false positives (when we loop unnecessarily)
but we can't tolerate false negatives (when oom_score_adj gets out of
sync). With the clone(CLONE_VM) race not addressed we are allowing
false negatives and IMHO that's not acceptable because it creates a
possibility for userspace to get an inconsistent picture. When
developing the patch I did think about using (p->mm->mm_users >
p->signal->nr_threads) condition and had to reject it due to that
reason.
>
> Oleg.
>
