Comment 6 for bug 1628336

Looking further into it, this is somehow caused by changes to the nbd module. Before 4.8 it was actually not creating any uevents on its own. Still there were events to be seen. With 4.8 they code changed to generate a change event, but also will delay the capacity change until the moment the socket is connected. And this seems to prevent whatever is otherwise producing the change event and triggering a partition scan. The nbd driver itself does not do that (like for example the loop driver does). So it could be that this had ever worked only because of the mysterious helper.

It seems that reverting a bit back to change the capacity as soon as it is set by ioctl (there are three ioctls which do this, so the values likely get touched multiple times, too) will again get the partition detected reliably. Not sure this can be guaranteed always...