A previous post about diskless Debian setup was using NFS. This is an attempt to try network block device. This can solve some problem caused by NFS, for example, running a docker daemon on such a system that uses overlay storage driver, which only supports a few backing file systems.

Network block device is a block device over network. On the server side, a daemon is running to deliver blocks and write the updates. On client side, a device is created in /dev and then we can mount it to another mount point. Hence from the user application’s perspective, it seems like a piece of hardware and we can format it to any file system.

The following documents the workflow from start to finish. In a very much similar way as the previous one. The workflow of the final system is:

1. Client boot with PXE stack on network card, which requests for DHCP
2. DHCP server respond with IP address and the location of the boot program
3. Client, according to the DHCP reply, request for the network boot program (NBP) from TFTP
4. Client pass the ownership to the NBP for the next stage of boot, which in case of pxelinux, will show boot menu and load kernel
5. The Linux kernel will work according to the parameters passed, which includes running nbd client and mount a nbd device as root.
6. Everything afterward is as usual, except all access to the “disk” is indeed going over the network to the nbd daemon.

## DHCP set up

In OpenWRT, the DHCP server is dnsmasq, which its configuration are located at /etc/config/dhcp (OpenWRT specific) and /etc/dnsmasq.conf (dnsmasq default). When we run it as a daemon, OpenWRT will create a new config at /var/etc that loads the latter and apply the attributes from the former. Hence it is better to modify at /etc/config/dhcp. Below should be a new section appended to the end:

config boot linux
option filename 'pxelinux.0'
option servername 'tftpserver'


The above will be converted into a line dhcp-boot=pxelinux.0,tftpserver,192.168.0.2 in the dnsmasq.conf, and delivered as DHCP option 66 (TFTP server address) and option 67 (boot filename) in the reply. In fact, we can make this more specific to one host, for example, the below will be host-based configuration based on MAC address:

config host 'myhost'
option ip '192.168.0.123'
option mac '01:23:45:ab:cd:ef'
option tag 'NOPXE'


## TFTP set up

TFTP is also supported in dnsmasq. What we needed to do is to add these two lines into /etc/dnsmasq.conf:

enable-tftp
tftp-root=/path/to/tftproot


or equivalently, add to /etc/config/dhcp:


config dnsmasq
# ...
option enable_tftp '1'
option tftp_root '/path/to/tftproot'

config dhcp 'lan'
option interface 'lan'
option dhcp_range '192.168.0.1,proxy'
option start '100'
option leasetime '12h'
option limit '150'
option dynamicdhcp '0'


The above configuration is for using a OpenWRT device to serve TFTP other than the one responding to DHCP. Otherwise, we do not need to modify the lan section to set the dhcp_range and dynamicdhcp option. But we must ensure that dnsmasq is listening to the interface that serves TFTP. It will not work if we ignored the interface for DHCP. If we set ths up in web interface in LuCI:

• in Network → Interface → LAN → DHCP server → General setup, uncheck “Ignore interface”
• in Network → Interface → LAN → DHCP server → Advanced settings, uncheck “Dynamic DHCP” and uncheck “Force”
• in Network → DHCP and DNS → General settings, uncheck “Authoritative”
• in Network → DHCP and DNS → TFTP settings, check “Enable TFTP server”, enter the full path for “TFTP server root”

To test, run “tftp <address>” and then try to “get " for the filename (or path) relative to the tftp root.

## Create NBD virtual disk

Conceptually, NBD is as simple as running a daemon and pass through all block device operations to a block device in hardware. But I don’t have a physical hard disk to spare so I use an image file instead.

On a separate computer, with the most updated Debian, run

apt-get install nbd-client


The installation of nbd-client will trigger initramfs-tools to create a nbd kernel module, which then will run update-initramfs to create /boot/initrd.img. This image, together with the corresponding kernel, will be copied to TFTP for PXE booting. Usually I will copy over the files /boot/initrd.img-5.10.0-18-amd64 and /boot/vmlinuz-5.10.0-18-amd64 to the tftproot and make symlinks initrd.img and vmlinuz to them, which the symlinks are referred in PXELinux config. Should I need more kernel modules for the diskless system at boot time, I do the similar: install it on another computer and rerun update-initramfs to generate a new initrd.img, then replace the one in TFTP and reboot the diskless system. If overwriting the initrd.img on the helper system is not desirable, use the following instead to create one in a different path:

mkinitramfs -d /etc/initramfs-tools -o path/to/initrd.img


Note: In my Debian installation of nbd-client, there is an issue on the initramfs script for using nbd as root. Two scripts are installed from nbd-client package, namely /usr/share/initramfs-tools/hooks/nbd and /usr/share/initramfs-tools/scripts/local-top/nbd. The latter is a script to reside in initrd.img and run at early phase of boot. The last few lines read:

# This should be removed once the cfq scheduler no longer deadlocks nbd
# devices
if grep -q '$cfq$' /sys/block/$nbdbasedev/queue/scheduler then echo deadline > /sys/block/$nbdbasedev/queue/scheduler
fi


which depends on grep. The former script ends with the line

copy_exec /sbin/nbd-client /sbin


which copies the nbd-client executable to /sbin. Booting from nbd will not success unless we remove the if block in the latter script or add another copy_exec to the former. Commenting out the if block is easier. Do this, and rerun update-initramfs -u.

Now, let’s create a disk of 8GB. To make an image of 8GB, we use truncate from coreutils or otherwise, using dd also works. Then format it, mount it, and install base Debian.

truncate -s 8G /server/FAKEDISK.img   # truncate is from coreutils
mkfs.ext4 /server/FAKEDISK.img
mount -t ext4 -o loop /server/FAKEDISK.img /mnt/point
debootstrap bullseye /mnt/point http://deb.debian.org/debian


Then we need some more basic set up on top of this barebone Debian. Since it will be mounted after PXE boot, the main network should already be set up (any driver required should be provided in initrd). But everything else is not. So we need to create user, set up SSH for remote access, allow sudo, and update hostname:

chroot /mnt/point
# now inside the chroot jail of a base debian system
apt-get update
apt-get -y install openssh-server sudo
echo <disklesshost> > /etc/hostname


Finally, unmount the loopback image and start to serve it

exit
# left the chroot jail
umount /mnt/point


## NBD server

The NBD server needs a daemon to serve:

apt-get install nbd-server


and this will create the user and group nbd. Since the daemon need to access to the image we created, we need:

chown nbd.nbd /server/FAKEDISK.img


The main cofig file for nbd server is /etc/nbd-server/config which will load everything in /etc/nbd-server/conf.d/*.conf. The man page for the config is man 5 nbd-server. We create one /etc/nbd-server/conf.d/diskless.conf:

[diskless]
exportname = /server/FAKEDISK.img
#authfile = /etc/nbd-server/allow


by default, port 10809 will be used but you can modify it in /etc/nbd-server/config. Here we export a name diskless and that links to the image file.

## PXElinux configuration

In the TFTP server, under the tftp root, we should make a copy of pxelinux.0 from syslinux and put the *.c32 file (e.g. 32-bit BIOS version) under boot/isolinux/. Then we can set up, for example, the below config at pxelinux.cfg/default or pxelinux.cfg/C0A80003 (address 192.168.0.3 in hex):

DEFAULT menu.c32

PROMPT 1
TIMEOUT 30
ONTIMEOUT Debian

LABEL reboot
COM32 reboot.c32

LABEL local
LOCALBOOT 0

LABEL Debian
KERNEL vmlinuz
APPEND vga=858 rw ip=dhcp initrd=initrd.img root=/dev/nbd0 nbdroot=192.168.0.2:diskless ipv6.disable=1


what to specify as kernel parameters is indeed platform-dependent and not found in nbd source code. Behind the scene in Debian, it is a shell script in initrd to parse them as command line parameters, which looks for nbdroot= and root=/dev/nbd*. Debian supports nbdroot=192.168.0.2,diskless,nbd0 as equivalent to above. Or we can put a port number as 192.168.0.2:10809/diskless. These parameters will translate into a nbd-client command (which resides in /sbin in initrd.img) to establish a connection.

Keep a copy of vmlinuz and initrd.img at the TFTP root. Start the nbd server at 192.168.0.2 as specified in the nbdroot option above. Then we are ready to boot.

## Final touch

The Debian system installed in this way will not come with kernel in the disk. Your /boot is empty but it is fine, until you need to load any additional kernel module and found modprobe failed. You may want to apt-get install linux-image-amd64.

Since the root is mounted by initramfs, you will have an empty /etc/fstab. You may want to remount the root with additional parameters, e.g.,

/dev/nbd0	/		ext4	remount,noatime,nodiratime	0	0


Using ext4 is not a requirement. See, for example https://wiki.archlinux.org/title/diskless_system, which uses btrfs instead of ext4 (but you will need to apt-get install some more packages to use it). It is reported that with compression support in btrfs, using nbd can be faster than nfs.