Configuration

nvme-cli and libnvme use a handful of files under /etc/nvme and /run/nvme to coordinate host identity, saved NVMe-oF connections, and shared state between independent tools on the same host.

NVMe-oF connection configuration

Which NVMe-oF controllers nvme connect-all / nvme discover should use is stored in an INI-format file, /etc/nvme/nvme-fabrics.conf, with an optional drop-in directory /etc/nvme/nvme-fabrics.conf.d/ for hosts that present multiple identities. This format, and the precedence rules for merging the main file with drop-ins, are documented in libnvme/design/CONFIG.md.

Older releases (pre-3.0) stored this configuration as /etc/nvme/config.json (JSON) or /etc/nvme/discovery.conf (argv-style). These formats are no longer read by libnvme directly; convert them once with:

$ nvme config-convert

See man nvme-config-convert for details.

Host identity

The default host NQN and host identifier are read from /etc/nvme/hostnqn and /etc/nvme/hostid. A host presenting multiple personalities can instead declare identity per drop-in file in nvme-fabrics.conf.d/ — see libnvme/design/CONFIG.md.

Coordinating multiple orchestrators on one host

NVMe-oF connections on a host are often managed by more than one independent actor at once — the initramfs (NBFT/FC-kickstart boot connections), a human running nvme connect-all / nvme disconnect-all, and daemons such as nvme-discoverd or nvme-stas. Two runtime mechanisms let them cooperate without stepping on each other:

• Ownership registry (/run/nvme/registry/): records which orchestrator owns each connected controller, so a sweeping command like nvme disconnect-all can avoid tearing down a connection another component depends on. This is per-boot runtime state written automatically by libnvme on connect. See libnvme/design/REGISTRY.md.
• Exclusion list (/etc/nvme/exclusions.conf, with an optional /etc/nvme/exclusions.conf.d/): a persistent, administrator-authored list of controllers that no orchestrator may auto-connect, useful for taking a path out of service (e.g. during maintenance or testing) regardless of which orchestrator would otherwise reconnect it. Manage it with the nvme exclusion-* commands or the nvme disconnect --exclude shortcut. See libnvme/design/EXCLUSIONS.md.

Both are cooperative mechanisms, not enforcement: every orchestrator runs as root and could bypass them, but well-behaved tools consult them first.