While the comment is good, let's actually enforce this with:

if (WARN_ON(!mutex_is_locked(&filter->notif_lock)))
	return NULL;

I don't see much use of lockdep in seccomp (well, any), but
wouldn't a stronger statement be to use lockdep, and just have:

lockdep_assert_held(&filter->notify_lock);

As that checks that the lock is held by the current task.
Although, that does put this check behind lockdep, which means
that running in "normal" circumstances is less safe (but faster?).