Skip to content

Booting from NVMe oF with NBFT

Daniel Wagner edited this page Nov 2, 2026 · 1 revision

Booting from NVMe-oF with NBFT

A host can boot from a namespace on an NVMe/TCP target. The UEFI firmware connects to the target and loads the boot loader from it. The firmware then describes the connection in an ACPI table, the NVMe Boot Firmware Table (NBFT). The operating system reads the NBFT and connects the same controllers again. The NBFT is defined in the NVM Express Boot Specification.

This page describes what nvme-cli does with the NBFT. The firmware setup depends on the vendor of the host.

Show the NBFT

The kernel provides the table as /sys/firmware/acpi/tables/NBFT (and NBFT1, ... if there are several). nvme nbft show decodes it:

# nvme nbft show

/sys/firmware/acpi/tables/NBFT:

NBFT Subsystems:

Idx|NQN                                                    |Trsp|Address      |SvcId|HFIs
---+-------------------------------------------------------+----+-------------+-----+----
1  |nqn.1988-11.com.dell:powerstore:00:88b402df2d762AA7AF94|tcp |172.18.240.60|4420 |1
2  |nqn.1988-11.com.dell:powerstore:00:88b402df2d762AA7AF94|tcp |172.18.240.60|4420 |1

NBFT HFIs:

Idx|Trsp|PCI Addr  |MAC Addr         |DHCP|IP Addr     |Mask|Gateway |DNS
---+----+----------+-----------------+----+------------+----+--------+--------
1  |tcp |0:6a:0.0  |00:62:0b:cb:eb:70|no  |172.18.240.1|24  |0.0.0.0 |0.0.0.0
  • Subsystems are the NVMe-oF subsystems to connect.
  • HFIs (host fabric interfaces) are the network interfaces of the host, with their address settings.
  • --discovery also shows the Discovery Controllers. -o json prints all details, including the host NQN.
  • --nbft-path=<dir> reads the tables from another directory, for example a copy for a bug report.

Connect the NBFT controllers

nvme connect-all --nbft connects the subsystems of the NBFT:

nvme connect-all --nbft

It connects through the network interface of the HFI. It records the owner nbft in the ownership registry. Because of this, nvme disconnect-all does not disconnect the boot volume. --owner=NAME records another owner.

In the initramfs

The initramfs must connect the boot namespace before it can mount the root file system. The initramfs generator does this, for example the nvmf module of dracut. It is not part of nvme-cli.

After the switch to the root file system

Some HFIs come up only after the initramfs, for example a second path for multipath. Then the host must connect the NBFT controllers again later. How this works depends on the nvme-cli version and build:

  • nvme-discoverd (built by default since nvme-cli 3.2): it reads the NBFT at start and connects its controllers. nbft = true in the [Global] section of nvme-discoverd.conf controls this. It is the default. It retries a failed connection, for example when an HFI comes up late. It works with any network management. See nvme-discoverd.
  • Legacy autoconnect (built when nvme-discoverd is not built): the NetworkManager dispatcher script 80-nvmf-connect-nbft.sh starts nvmf-connect-nbft.service when an NBFT interface comes up. The service runs nvme connect-all --nbft. There is no such hook for systemd-networkd. On a host without NetworkManager, a late HFI is not connected.

Network interface names

The initramfs names the network interfaces of the HFIs nbft0, nbft1, and so on. The udev rule 65-persistent-net-nbft.rules keeps these names after the switch to the root file system, so that udev does not rename them. nvme-cli installs this rule with every NVMe-oF build.

Take an NBFT path out of service

The NBFT is in the firmware. You cannot edit it on the host. To stop the host from connecting an NBFT controller, add it to the exclusion list:

nvme exclusion add --entry="nqn=<subsystem NQN>;traddr=<address>"

nvme connect-all --nbft and nvme-discoverd then skip it.

See also

Clone this wiki locally