nvme-discoverd(8) Manual Page

nvme-discoverd is a persistent daemon that connects the host’s desired NVMe-oF controllers and keeps them connected. It reads three sources to build its set of desired connections: the NBFT ACPI table (boot-time controllers), the shared NVMe-oF fabrics configuration (nvme-fabrics.conf(5)), and the Discovery Log Pages of any Discovery Controllers it connects to (including FC targets found via FC kickstart). It never disconnects a live controller in response to a discovery change; the only disconnects it causes are a side effect of stopping its own systemd units, at shutdown or on request.

At shutdown, a unit is stopped only after the file systems mounted with the _netdev option are unmounted. A file system on an NVMe-oF namespace must use _netdev in fstab(5). Without it, the controller can be disconnected while the file system is still mounted.

Every connection attempt is delegated to nvme connect, which runs inside a systemd transient unit instead of writing directly to /dev/nvme-fabrics. As a result, a stuck connect attempt never blocks the daemon’s event loop. Each unit is named after a hash of the controller’s transport identity.

Failed connections retry with exponential backoff. Discovery Controllers from NBFT or the fabrics config represent explicit intent, so they always retry forever. A Discovery Controller found only through a referral or FC kickstart is different: it gives up after dc-giveup-timeout (default 72hours) of unbroken failure and is dropped from tracking.

A TLS connection needs tlshd(8) from ktls-utils. Enable it with systemctl enable --now tlshd. When both are enabled, nvme-discoverd starts after tlshd.

Before connecting or reconnecting anything, nvme-discoverd checks the system-wide exclusion list. It also registers ownership of every controller it connects (--owner discoverd, or --owner nbft for an NBFT-sourced controller) in the NVMe-oF ownership registry. See nvme-registry-list(1) and nvme-exclusion-add(1).

A Discovery Controller that has no owner, for example one connected with nvme discover --persistent, is not tracked. When its discovery log changes, or when it reconnects, nvme-discoverd runs nvme connect-all for it in a transient unit, as the legacy udev rules do. The connections it makes have no owner, and nvme disconnect-all removes them. Events from a Discovery Controller with another owner are ignored.

nvme-discoverd is still experimental. It is meant to replace the older udev-rule/systemd-unit autoconnect mechanism. By default, the legacy mechanism is built only when nvme-discoverd is not. Building with -Dnvmf-autoconnect=enabled installs both. The legacy udev rules are active as soon as they are installed. nvme-discoverd runs only when the administrator enables it with systemctl enable --now nvme-discoverd. While nvme-discoverd is active, nvmf-autoconnect.service and the FC kickstart skip their connect step.

nvme-discoverd.conf holds the daemon’s own knobs, in a [Global] section and a [Discovery] section:

nbft (boolean, default true) adopts and connects controllers listed in the NBFT ACPI table. debug-level (err, warn, info, or debug; default info) sets the log threshold; --debug on the command line overrides it.

epcsd-poll-interval-minutes (default 15, minimum 1) sets how often, in minutes, to reconnect and re-check a DC whose effective EPCSD (Explicit Persistent Connection Support for Discovery) is 0. Such a DC is disconnected between checks. fc-kickstart-interval-minutes (default 0) re-issues FC kickstart every N minutes; 0 = only at startup and on every FC controller drop. dc-giveup-timeout (default 72hours) is the give-up horizon described above. It takes a systemd time span ("72hours", "3 days 5 hours") or a unit-less value in seconds; "infinity" = never give up, 0 = give up on the first failure.

A DC connects with a keep-alive timeout of 30 seconds, or with its configured keep-alive-tmo (see nvme-fabrics.conf(5)). A DC configured with persistent = no connects with none.

zeroconf (boolean, default false) connects to and manages Discovery Controllers found through mDNS (TP8009). It requires systemd-resolved v258 or later with mDNS enabled (MulticastDNS=yes), and an nvme-discoverd built with mDNS support. It is re-read on SIGHUP.

The connections nvme-discoverd manages come entirely from the shared fabrics configuration, not from this file: which Discovery Controllers and subsystems to connect, host identity, per-connection parameters. See nvme-fabrics.conf(5).

nvme-fabrics.conf(5) tlshd(8) nvme-connect(1) nvme-registry-list(1) nvme-exclusion-add(1)

Part of the nvme-user suite