Diskless system

Contrast to the official installation guide, which concentrates on an arch installation on local storage, here the installation is to be placed on network storage.

Server configuration

You will need a DHCP server to setup networking, as well as a TFTP server to transfer the boot image (a requirement of all PXE option roms). Unlike PXE which is designed for quick-and-dirty/temporary setups to boot installation media, here you're actually doing a network installation, which is more-permanent, and thus calls for doing things properly.

DHCP

Install ISC dhcp.

# pacman -Syu dhcp

Configure ISC DHCP.

# vim /etc/dhcpd.conf

allow booting;
allow bootp;

authoritative;

option domain-name-servers 10.0.0.1;

group 
{
    next-server 10.0.0.1;
    filename "pxelinux.0";

    subnet 10.0.0.0 netmask 255.255.255.0 {
        option routers 10.0.0.1;
        range 10.0.0.128 10.0.0.254;
    }
}

Note: next-server should be the address of the TFTP server; everything else should be changed to match your network

Start ISC DHCP.

# systemctl start dhcpd4

TFTP

The TFTP server will be used to transfer the kernel, initramfs, and pxelinux to the client.

Install tftp-hpa.

# pacman -Syu tftp-hpa

Note: Change /srv/arch/boot to /srv/boot if you're going to use NBD, because you'll be unable to mount your root filesystem while in use

Configuration

Copy the tftp unit files in /usr/lib/systemd/system to /etc/systemd/system; the former gets overwritten when systemd is updated. The systemd article talks in more detail about customizing unit files.

# cp /usr/lib/systemd/system/tftpd.s* /etc/systemd/system/

Next we tell tftpd where the tftp root is: /srv/arch/boot.

# vim /etc/systemd/system/tftpd.service

[Unit]
Description=hpa's original TFTP daemon

[Service]
ExecStart=/usr/sbin/in.tftpd -s /srv/arch/boot/
StandardInput=socket
StandardOutput=inherit
StandardError=journal

Start tftpd :

# systemctl start tftpd.socket

Network storage

The primary difference between using NFS and NBD is while with both you can in fact have multiple clients using the same installation, with NBD (by the nature of manipulating a filesystem directly) you'll need to use the copyonwrite mode to do so, which ends up discarding all writes on client disconnect. In some situations however, this might be highly desirable.

NFS

Install nfs-utils on the server.

# pacman -Syu nfs-utils

You'll need to add the root of your arch installation to your NFS exports:

# vim /etc/exports

/srv/arch *(rw,fsid=0,no_root_squash,no_subtree_check,async)

Note: If you're worried about data loss in the event of network failure, replace async with sync--additional options can be found in the NFS article.

Next start NFS.

# systemctl start rpc-idmapd.service rpc-mountd.service

NBD

Install nbd.

# pacman -Syu nbd

Configure nbd.

# vim /etc/nbd-server/config

[generic]
[arch]
    exportname = /srv/arch.img
    copyonwrite = false

Note: Set copyonwrite to true if you want to have multiple clients using the same NBD share simultaneously; refer to man 5 nbd-server for more details.

Start nbd.

# systemctl start nbd

Client installation

Directory setup

Create a sparse file of at least 1 gigabyte, and create an ext4 filesystem on it (you can of course also use a real block device or LVM if you so desire).

# truncate -s 1G /srv/arch.img
# mkfs.ext4 /srv/arch.img
# export root=/srv/arch
# mkdir -p "$root"
# mount /srv/arch.img "$root"

Note: Creating a separate filesystem is required for NBD but optional for NFS and can be skipped/ignored.

Bootstrapping installation

Install devtools and arch-install-scripts, and run mkarchoot.

# pacman -Syu devtools arch-install-scripts
# mkarchroot -f --arch x86_64 "$root" base mkinitcpio-nfs-utils nfs-utils

Replace x86_64 with i686 as appropriate for your target hardware.

Note: If you're using NBD, you'll want to replace nfs-utils with mkinitcpio-nbd^AUR. mkinitcpio-nfs-utils is still required--ipconfig used in early-boot is provided only by the latter.

Trivial modifications to the net hook are required in order for NFSv4 mounting to work (not supported by nfsmount--the default for the net hook).

# sed s/nfsmount/mount.nfs4/ "$root/usr/lib/initcpio/hooks/net" | tee "$root/usr/lib/initcpio/hooks/net_nfs4"
# cp "$root/usr/lib/initcpio/install/{net,net_nfs4}"

The copy of net is unfortunately needed so it does not get overwritten when mkinitcpio-nfs-utils is updated on the client installation.

The newly-created net_nfs4 hook needs to be added to the HOOKS array in mkinitcpio.conf, and mount.nfs4 to BINARIES.

# vim "$root/etc/mkinitcpio.conf"

MODULES="nfsv4"
HOOKS="base udev autodetect net_nfs4 filesystems"
BINARIES="/sbin/mount.nfs4"

Note: The upstream net hook is sufficient for NBD installations; you will also need to append nbd to your HOOKS array; mount.nfs4 is also not needed

The initramfs now needs to be rebuilt; the easiest way to do this is via arch-chroot.

# arch-chroot "$root" /bin/bash
(chroot) # mkinitcpio -p linux
(chroot) # exit

Client configuration

In addition to the setup mentioned here, you should also set up your hostname, timezone, locale, and keymap, and follow any other relevant parts of the Installation Guide.

Bootloader

Pxelinux

Install syslinux.

# pacman -Syu syslinux

Copy the pxelinux bootloader (provided by the syslinux package) to the boot directory of the client.

# cp /usr/lib/syslinux/pxelinux.0 "$root/boot"
# mkdir "$root/boot/pxelinux.cfg"

We also created the pxelinux.cfg directory, which is where pxelinux searches for configuration files by default. Because we don't want to discriminate between different host MACs, we then create the default configuration.

# vim "$root/boot/pxelinux.cfg/default"

default linux

label linux
kernel vmlinuz-linux
append initrd=initramfs-linux.img ip=:::::eth0:dhcp nfsroot=10.0.0.1:/

Or if you are using NBD, use the following append line:

append initrd=initramfs-linux.img ip=:::::eth0:dhcp nbd_host=10.0.0.1 nbd_name=arch

Note: You will need to change nbd_host and/or nfsroot, respectively, to match your network configuration (the address of the NFS/NBD server)

The pxelinux configuration syntax identical to syslinux; refer to the upstream documentation for more information.

The kernel and initramfs will be transferred via TFTP, so the paths to those are going to be relative to the TFTP root. Otherwise, the root filesystem is going to be the NFS mount itself, so those are relative to the root of the NFS server.

GRUB

Though poorly documented, grub also supports being loaded via PXE. Because of an endainness bug in grub-core/net/tftp.c not fixed until bzr revision 4548 (grub-bios 2.00 is revision 4542), you will need to install grub-bios-bzr^AUR; it makes the most sense to do this on the client installation. The Arch User Repository article describes in more detail on how to build AUR packages.

Note: If you build grub-bios-bzr^AUR on the host installation and your target architecture doesn't match; you'll end up needing to build it on the target installation instead

# pacman --root "$root" --dbpath "$root/var/lib/pacman" -U grub-bios-bzr-4599-1-x86_64.pkg.tar.xz

Only because we want to use the grub-bios-bzr^AUR installation on the target, we arch-chroot the target installation so we can use its grub-mknetdir instead.

# arch-chroot /srv/arch
(chroot) # grub-mknetdir --net-directory=/boot/ --subdir=/grub
(chroot) # exit

Note: You will need to change the filename directive in /etc/dhcpd.conf to /grub/i386-pc/core.0 as grub-mknetdir will tell you

Now we create a trivial QnD grub configuration

# vim "$root/boot/grub/grub.cfg"

menuentry "Arch Linux" {
    linux /vmlinuz-linux ip=:::::eth0:dhcp nfsroot=10.0.0.1:/
    initrd /initramfs-linux.img
}

GRUB dark-magic will set root=(pxe:10.0.0.1) automatically, so that the kernel and initramfs are transferred via TFTP without any additional configuration.

Note: Modify your kernel line as-necessary, refer to #Pxelinux for NBD-related options

VFS mountpoints

Add hacks to your fstab for the root filesystem and devpts.

# vim "$root/etc/fstab"

none    /           none      defaults                        0 0
none    /dev/pts    devpts    mode=0620,gid=5,nosuid,noexec   0 0

Consider /var

For a simple setup, it may not be an issue that /var is shared between all of the clients. A booted system does not expect /var to be shared. If you intend to run BOINC or SLURM^AUR, you will definitely need to have separate /var directories for each of your clients.

You could mount /var/log, for example, as tmpfs so that logs from multiple hosts don't mix unpredictably.

/etc/fstab

tmpfs   /var/log    tmpfs     nodev,nosuid                    0 0

Client boot

NBD

If you're using NBD, you'll need to umount the arch.img before/while you boot your client.

This makes things particularly interesting when it comes to kernel updates. You can't have your client filesystem mounted while you're booting a client, but that also means you need to use a kernel separate from your client filesystem in order to build it.

You'll need to first copy $root/boot from the client installation to your tftp root (i.e. /srv/boot).

# cp -r "$root/boot" /srv/boot

You'll then need to umount $root before you start the client.

# umount "$root"

Note: To update the kernel in this setup, you either need to mount /srv/boot using NFS in fstab on the client (prior to doing the kernel update) or mount your client filesystem after the client has disconnected from NBD

References

kernel.org: Mounting the root filesystem via NFS (nfsroot)

syslinux.org: pxelinux FAQ