On Wed, Jul 06, 2022 at 09:42:19AM +0200, Christoph Hellwig wrote: > The mdev_type already holds a reference to the parent through > mdev_types_kset, so drop the extra reference. I would drop this patch, but at least the explanation needs tweaking.. kobj's are weird things, they have a kobj.parent, but the refcount for that is dropped during kobject_del() and parent is NULL'd. vs this reference which is being put back at kobject release. So, the extra kset reference isn't contributing more protection - when mdev_unregister_device() calls kset_unregister() it still has a valid reference on the parent from the caller. If this change is safe it is because no accesses to mdev->parent are happening after mdev_unregister_device() returns, due to things like the sysfs attribute fencing and the mdev_unreg_sem fencing child devices. I didn't check everything, but it seems believable it could be true. I also suggest this patch should NULL mdev_type->parent during unregister_device as the kobj stuff does, so any use after unregister but before release could be crash, we generally shouldn't leave unref'd pointers floating around. Jason
