#!/usr/bin/env bash # Installs NixOS on a Hetzner server, wiping the server. # # This is for a specific server configuration; adjust where needed. # # # Usage: # ssh root@YOUR_SERVERS_IP bash -s < hetzner-dedicated-wipe-and-install-nixos.sh # # When the script is done, make sure to boot the server from HD, not rescue mode again. # Explanations: # # * Following largely https://nixos.org/nixos/manual/index.html#sec-installing-from-other-distro. # * and https://nixos.wiki/wiki/NixOS_on_ZFS # * **Important:** First you need to boot in legacy-BIOS mode. Then ask for # hetzner support to enable UEFI for you. # * We set a custom `configuration.nix` so that we can connect to the machine afterwards, # inspired by https://nixos.wiki/wiki/Install_NixOS_on_Hetzner_Online # * This server has 2 SSDs. # We put everything on mirror (RAID1 equivalent). # * A root user with a password is created, so that you can just login # as root when using the Hetzner spider KVM. # * The script reboots at the end. # * exports of env vars are added throughout the script in case you want to run it manually export LC_ALL=C # WARNING: on 2023/07/16 the rescue system of hetzner boots with kernel 6.3.7 which # is by default not supported by the latest debian package. You need to update to debian # unstable to proceed with the zfs installation. cat > /etc/apt/preferences.d/90_zfs <` because the file may already contain some detected RAID arrays, # which would take precedence over our ``. echo 'AUTO -all ARRAY UUID=00000000:00000000:00000000:00000000' > /etc/mdadm/mdadm.conf # Create wrapper for parted >= 3.3 that does not exit 1 when it cannot inform # the kernel of partitions changing (we use partprobe for that). echo -e "#! /usr/bin/env bash\nset -e\n" 'parted $@ 2> parted-stderr.txt || grep "unable to inform the kernel of the change" parted-stderr.txt && echo "This is expected, continuing" || echo >&2 "Parted failed; stderr: $(< parted-stderr.txt)"' > parted-ignoring-partprobe-error.sh && chmod +x parted-ignoring-partprobe-error.sh # Create partition tables (--script to not ask) ./parted-ignoring-partprobe-error.sh --script $DISK1 mklabel gpt ./parted-ignoring-partprobe-error.sh --script $DISK2 mklabel gpt # Create partitions (--script to not ask) # # We create the 1MB BIOS boot partition at the front. # # Note we use "MB" instead of "MiB" because otherwise `--align optimal` has no effect; # as per documentation https://www.gnu.org/software/parted/manual/html_node/unit.html#unit: # > Note that as of parted-2.4, when you specify start and/or end values using IEC # > binary units like "MiB", "GiB", "TiB", etc., parted treats those values as exact # # Note: When using `mkpart` on GPT, as per # https://www.gnu.org/software/parted/manual/html_node/mkpart.html#mkpart # the first argument to `mkpart` is not a `part-type`, but the GPT partition name: # ... part-type is one of 'primary', 'extended' or 'logical', and may be specified only with 'msdos' or 'dvh' partition tables. # A name must be specified for a 'gpt' partition table. # GPT partition names are limited to 36 UTF-16 chars, see https://en.wikipedia.org/wiki/GUID_Partition_Table#Partition_entries_(LBA_2-33). # TODO the bios partition should not be this big # however if it's less the installation fails with # cannot copy /nix/store/d4xbrrailkn179cdp90v4m57mqd73hvh-linux-5.4.100/bzImage to /boot/kernels/d4xbrrailkn179cdp90v4m57mqd73hvh-linux-5.4.100-bzImage.tmp: No space left on device ./parted-ignoring-partprobe-error.sh --script --align optimal $DISK1 -- mklabel gpt \ mkpart 'BIOS-boot-partition' 1MB 2MB set 1 bios_grub on \ mkpart 'EFI-system-partition' 2MB 512MB set 2 esp on \ mkpart 'data-partition' 512MB '100%' ./parted-ignoring-partprobe-error.sh --script --align optimal $DISK2 -- mklabel gpt \ mkpart 'BIOS-boot-partition' 1MB 2MB set 1 bios_grub on \ mkpart 'EFI-system-partition' 2MB 512MB set 2 esp on \ mkpart 'data-partition' 512MB '100%' # Reload partitions partprobe # Wait for all devices to exist udevadm settle --timeout=5 --exit-if-exists=$DISK1-part1 udevadm settle --timeout=5 --exit-if-exists=$DISK1-part2 udevadm settle --timeout=5 --exit-if-exists=$DISK1-part3 udevadm settle --timeout=5 --exit-if-exists=$DISK2-part1 udevadm settle --timeout=5 --exit-if-exists=$DISK2-part2 udevadm settle --timeout=5 --exit-if-exists=$DISK2-part3 # Wipe any previous RAID signatures # sometimes they are not on a specific disk for some reason mdadm --zero-superblock --force $DISK1-part1 || true mdadm --zero-superblock --force $DISK1-part2 || true mdadm --zero-superblock --force $DISK1-part3 || true mdadm --zero-superblock --force $DISK2-part1 || true mdadm --zero-superblock --force $DISK2-part2 || true mdadm --zero-superblock --force $DISK2-part3 || true # Creating file systems changes their UUIDs. # Trigger udev so that the entries in /dev/disk/by-uuid get refreshed. # `nixos-generate-config` depends on those being up-to-date. # See https://github.com/NixOS/nixpkgs/issues/62444 udevadm trigger # Sometimes the zfs module is not loaded modprobe zfs # taken from https://nixos.wiki/wiki/NixOS_on_ZFS # somehow there is a weird symlink in the default zfs zpool create -O mountpoint=none \ -O atime=off \ -O compression=lz4 \ -O xattr=sa \ -O acltype=posixacl \ -o ashift=12 \ -f \ root_pool mirror $DISK1-part3 $DISK2-part3 zfs create -o mountpoint=legacy \ -o encryption=on \ -o keyformat=passphrase \ root_pool/encrypted zfs create -o mountpoint=legacy root_pool/encrypted/root # add 1G of reseved space in case the disk gets full # zfs needs space to delete files zfs create -o refreservation=1G -o mountpoint=none root_pool/reserved # all application data goes here. Only backups from this dataset and its # children are important zfs create -o mountpoint=legacy root_pool/encrypted/data # this creates a special volume for db data see https://wiki.archlinux.org/index.php/ZFS#Databases zfs create -o mountpoint=legacy \ -o recordsize=8K \ -o primarycache=metadata \ -o logbias=throughput \ root_pool/encrypted/data/postgresql # NixOS pre-installation mounts # # Mount the filesystems manually. The nixos installer will detect these mountpoints # and save them to /mnt/nixos/hardware-configuration.nix during the install process. mount -t zfs root_pool/root /mnt mkdir -p /mnt/var/lib/postgresql mount -t zfs root_pool/postgresql /mnt/var/lib/postgresql # Create a raid mirror for the efi boot # see https://docs.hetzner.com/robot/dedicated-server/operating-systems/efi-system-partition/ # TODO check this though the following article says it doesn't work properly # https://outflux.net/blog/archives/2018/04/19/uefi-booting-and-raid1/ mdadm --create --run --verbose /dev/md127 \ --level 1 \ --raid-disks 2 \ --metadata 1.0 \ --homehost=$MY_HOSTNAME \ --name=boot_efi \ $DISK1-part2 $DISK2-part2 # Assembling the RAID can result in auto-activation of previously-existing LVM # groups, preventing the RAID block device wiping below with # `Device or resource busy`. So disable all VGs first. vgchange -an # Wipe filesystem signatures that might be on the RAID from some # possibly existing older use of the disks (RAID creation does not do that). # See https://serverfault.com/questions/911370/why-does-mdadm-zero-superblock-preserve-file-system-information wipefs -a /dev/md127 # Disable RAID recovery. We don't want this to slow down machine provisioning # in the rescue mode. It can run in normal operation after reboot. echo 0 > /proc/sys/dev/raid/speed_limit_max # Filesystems (-F to not ask on preexisting FS) mkfs.vfat -F 32 /dev/md127 # Creating file systems changes their UUIDs. # Trigger udev so that the entries in /dev/disk/by-uuid get refreshed. # `nixos-generate-config` depends on those being up-to-date. # See https://github.com/NixOS/nixpkgs/issues/62444 udevadm trigger mkdir -p /mnt/boot/efi mount /dev/md127 /mnt/boot/efi # Installing nix # Allow installing nix as root, see # https://github.com/NixOS/nix/issues/936#issuecomment-475795730 mkdir -p /etc/nix echo "build-users-group =" > /etc/nix/nix.conf # using determinate systems installer, for more information # check https://github.com/DeterminateSystems/nix-installer curl --proto '=https' --tlsv1.2 -sSf -L https://install.determinate.systems/nix | sh -s -- install # Keep in sync with `system.stateVersion` set below! nix-channel --add https://nixos.org/channels/nixos-23.05 nixpkgs nix-channel --update # TODO use something like nix shell nixpkgs#nixos-generate-config nixpkgs#nixos-install nixpkgs#nixos-enter nixpkgs#manual.manpages # Getting NixOS installation tools nix-env -iE "_: with import { configuration = {}; }; with config.system.build; [ nixos-generate-config nixos-install nixos-enter manual.manpages ]" # TODO # perl: warning: Please check that your locale settings: # LANGUAGE = (unset), # LC_ALL = "en_US.UTF-8", # LANG = "en_US.UTF-8" # are supported and installed on your system. nixos-generate-config --root /mnt # Find the name of the network interface that connects us to the Internet. # Inspired by https://unix.stackexchange.com/questions/14961/how-to-find-out-which-interface-am-i-using-for-connecting-to-the-internet/302613#302613 export RESCUE_INTERFACE=$(ip route get 8.8.8.8 | grep -Po '(?<=dev )(\S+)') # Find what its name will be under NixOS, which uses stable interface names. # See https://major.io/2015/08/21/understanding-systemds-predictable-network-device-names/#comment-545626 # NICs for most Hetzner servers are not onboard, which is why we use # `ID_NET_NAME_PATH`otherwise it would be `ID_NET_NAME_ONBOARD`. export INTERFACE_DEVICE_PATH=$(udevadm info -e | grep -Po "(?<=^P: )(.*${RESCUE_INTERFACE})") export UDEVADM_PROPERTIES_FOR_INTERFACE=$(udevadm info --query=property "--path=$INTERFACE_DEVICE_PATH") export NIXOS_INTERFACE=$(echo "$UDEVADM_PROPERTIES_FOR_INTERFACE" | grep -o -E 'ID_NET_NAME_PATH=\w+' | cut -d= -f2) echo "Determined NIXOS_INTERFACE as '$NIXOS_INTERFACE'" export IP_V4=$(ip route get 8.8.8.8 | grep -Po '(?<=src )(\S+)') echo "Determined IP_V4 as $IP_V4" # Find what its name will be under NixOS, which uses stable interface names. # See https://major.io/2015/08/21/understanding-systemds-predictable-network-device-names/#comment-545626 # NICs for most Hetzner servers are not onboard, which is why we use # `ID_NET_NAME_PATH`otherwise it would be `ID_NET_NAME_ONBOARD`. export INTERFACE_DEVICE_PATH=$(udevadm info -e | grep -Po "(?<=^P: )(.*${RESCUE_INTERFACE})") export UDEVADM_PROPERTIES_FOR_INTERFACE=$(udevadm info --query=property "--path=$INTERFACE_DEVICE_PATH") export NIXOS_INTERFACE=$(echo "$UDEVADM_PROPERTIES_FOR_INTERFACE" | grep -o -E 'ID_NET_NAME_PATH=\w+' | cut -d= -f2) echo "Determined NIXOS_INTERFACE as '$NIXOS_INTERFACE'" # Determine Internet IPv6 by checking route, and using ::1 # (because Hetzner rescue mode uses ::2 by default). # The `ip -6 route get` output on Hetzner looks like: # # ip -6 route get 2001:4860:4860:0:0:0:0:8888 # 2001:4860:4860::8888 via fe80::1 dev eth0 src 2a01:4f8:151:62aa::2 metric 1024 pref medium export IP_V6="$(ip route get 2001:4860:4860::8888 | head -1 | cut -d' ' -f7 | cut -d: -f1-4)::1" echo "Determined IP_V6 as $IP_V6" # From https://stackoverflow.com/questions/1204629/how-do-i-get-the-default-gateway-in-linux-given-the-destination/15973156#15973156 read _ _ DEFAULT_GATEWAY _ < <(ip route list match 0/0); echo "$DEFAULT_GATEWAY" echo "Determined DEFAULT_GATEWAY as $DEFAULT_GATEWAY" # Generate `configuration.nix`. Note that we splice in shell variables. cat > /mnt/etc/nixos/configuration.nix <