Re: [PATCH 4/6] mm, oom: skip vforked tasks from being selected

[Oleg Nesterov <oleg@redhat.com>]

On 05/31, Michal Hocko wrote:
>
> On Mon 30-05-16 21:28:57, Oleg Nesterov wrote:
> >
> > I don't think we can trust vfork_done != NULL.
> >
> > copy_process() doesn't disallow CLONE_VFORK without CLONE_VM, so with this patch
> > it would be trivial to make the exploit which hides a memory hog from oom-killer.
>
> OK, I wasn't aware of this possibility.

Neither was me ;) I noticed this during this review.

> > Or I am totally confused?
>
> I cannot judge I am afraid. You are definitely much more familiar with
> all these subtle details than me.

OK, I just verified that clone(CLONE_VFORK|SIGCHLD) really works to be sure.

> +/* expects to be called with task_lock held */
> +static inline bool in_vfork(struct task_struct *tsk)
> +{
> +	bool ret;
> +
> +	/*
> +	 * need RCU to access ->real_parent if CLONE_VM was used along with
> +	 * CLONE_PARENT
> +	 */
> +	rcu_read_lock();
> +	ret = tsk->vfork_done && tsk->real_parent->mm == tsk->mm;
> +	rcu_read_unlock();
> +
> +	return ret;
> +}

Yes, but may I ask to add a comment? And note that "expects to be called with
task_lock held" looks misleading, we do not need the "stable" tsk->vfork_done
since we only need to check if it is NULL or not.

It would be nice to explain that

	1. we check real_parent->mm == tsk->mm because CLONE_VFORK does not
	   imply CLONE_VM

	2. CLONE_VFORK can be used with CLONE_PARENT/CLONE_THREAD and thus
	   ->real_parent is not necessarily the task doing vfork(), so in
	   theory we can't rely on task_lock() if we want to dereference it.

	   And in this case we can't trust the real_parent->mm == tsk->mm
	   check, it can be false negative. But we do not care, if init or
	   another oom-unkillable task does this it should blame itself.

Oleg.