> but there are several cases where this is undesirable so I do not > think the kernel should do it. > Having userspace handle policy decisions allows for more flexibility. Ewan, could you elaborate on this point? A volume ACL removal on the target is not a transient disruption to access to the volume - it is permanent and deliberate. Keeping the device data structures around on the host paints a false picture for applications, as if those devices are still accessible. Moreover, if a new volume is connected with the same LUN, the old device nodes are re-used with no indication to the application that the underlying volume has changed. As Brian showed above, this behavior can cause corruption when devices are accessed via multipath. Sure, this can be avoided via a manual rescan-scsi-bus.sh -r after removing volume ACLs. But we already have automatic scanning of the target upon receiving the REPORTED_LUNS_DATA_HAS_CHANGED UA, and it seems unnecessarily asymmetric for this scanning to have the ability to create new devices but not delete old ones. As far as policy decisions go, NVMe has in-kernel scanning which can both add and remove devices. Is it a protocol difference that prevents SCSI from doing the same? Thanks, Uday
