On Sat, 2017-02-18 at 21:16 +0530, Sriram Popuri wrote: The NVMe-oF target currently has two distinct modes for its subsystems - Direct and Virtual. We're working to unify these and make life a bit simpler for everyone, but for now we'll need to treat them separately. For direct mode, the subsystem exported is identical to a physical NVMe subsystem attached to the target system. We're just exporting a physical device over the network exactly as it is, with the absolute minimum amount of virtualization required by the NVMe-oF specification. All commands from the initiator simply pass through the NVMe-oF layer and are forwarded directly to the backing SSD. In this case, adding namespaces dynamically requires the backing NVMe SSD to support dynamic namespaces. If the SSD supports it, then it should[1] work today. For virtual mode, the NVMe-oF target is exporting a virtualized NVMe subsystem that may be backed by any type of device. This requires our target to emulate lots of functionality that the backing device may or may not actually support. Adding namespaces through the config file is only for virtual mode. We could of course emulate all of namespace management in virtual mode, and we definitely should, but the code to do this has not been implemented yet. 1. AER isn't emulated in virtual mode today, which is a problem. Even in direct mode AER isn't fully set up today, although we have someone working on this part now.