I have looked over the patch a bit more and I think there is an issue in it as it decouples the epoch update formerly in __sd_leave_done from the vnode information update the line below - it seems like we might now hit a window where the vnode has alreayd been incremented by operations still operate on the old vnode confifuration. But at least to me it seems like we could update the vnode configuration earlier, too. another thing I noticed is that I think your patch makes the ->dispatch cluster driver operation obsolete. Once we never register the event handler for the cluster driver FD it can be handled netirely inside the cluster driver, e.g. just call register_event from inside the init method and don't even tell the core about it. |