On Thu 02-06-16 21:20:03, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Wed 01-06-16 15:53:13, Andrew Morton wrote:
[...]
> > > Is it even possible to hit that race?
> >
> > It is, we can have a concurrent mmput followed by mmdrop.
> >
> > > find_lock_task_mm() takes some
> > > care to prevent a NULL ->mm. But I guess a concurrent mmput() doesn't
> > > require task_lock(). Kinda makes me wonder what's the point in even
> > > having find_lock_task_mm() if its guarantee on ->mm is useless...
> >
> > find_lock_task_mm makes sure that the mm stays non-NULL while we hold
> > the lock. We have to do all the necessary pinning while holding it.
> > atomic_inc_not_zero will guarantee we are not racing with the finall
> > mmput.
> >
> > Does that make more sense now?
>
> what Andrew wanted to confirm is "how can it be possible that
> mm->mm_users < 1 when there is a tsk with tsk->mm != NULL", isn't it?
>
> Indeed, find_lock_task_mm() returns a tsk where tsk->mm != NULL with
> tsk->alloc_lock held. Therefore, tsk->mm != NULL implies mm->mm_users > 0
> until we release tsk->alloc_lock , and we can do
>
> p = find_lock_task_mm(tsk);
> if (!p)
> goto unlock_oom;
>
> mm = p->mm;
> - if (!atomic_inc_not_zero(&mm->mm_users)) {
> - task_unlock(p);
> - goto unlock_oom;
> - }
> + atomic_inc(&mm->mm_users);
>
> task_unlock(p);
>
> in __oom_reap_task() (unless I'm missing something).
OK, I guess you are right. Care to send a patch? That also means that
your patch to set mm = NULL in the atomic_inc_not_zero path is not
really needed and in fact e2fe14564d3316d1625ed20bf1083995f4960893 which
is sitting in the Linus tree is OK.
I will comment on the rest in a separate reply to not mix the two
things.
--
Michal Hocko
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>