nixos/docs/proxmox-images.md

Proxmox VM disk images

proxmox-* hosts (VM platform, not lxc-*) can be built as standalone, ready-to-attach .raw disk images via disko's own image-builder — no nixos-install, no live installer boot. This uses the same disko.devices config (modules/disko/proxmox.nix) already used to format a real disk on install, so there's nothing host-specific to write; it's available for every proxmox-* target automatically.

scripts/proxmox/create-proxmox-resource.sh --type vm --host <name> automates the whole walkthrough below (and the equivalent LXC one) end to end, including host-key handling and building the image directly on the Proxmox node itself (no local build, no image transfer) — see its --help. The steps here are what it runs under the hood, useful for doing any of it by hand or understanding what it does before you trust it against real infrastructure.

Building

nix build .#nixosConfigurations.proxmox-server.config.system.build.diskoImagesScript
sudo ./result --build-memory 2048

This produces <hostname>.raw in the current directory (e.g. server.raw for proxmox-server, matching networking.hostName, not the flake attribute name — every proxmox-* host gets a distinctly named image instead of all of them producing an identical main.raw). The script builds inside a temporary QEMU VM and moves the finished image out to the working directory when done; --build-memory controls how much RAM that build VM gets.

disko.devices.disk.main.imageSize (currently 20G, in modules/disko/proxmox.nix) sets the image's total size — disko doesn't support auto-resizing, so this needs to comfortably fit ESP + swap + root at build time. Grow the virtual disk (and resize the filesystem) in Proxmox after attaching if a host needs more than that; this is the normal way to size these images, not a one-time decision to get exactly right up front.

Host keys

The disko image script runs a real activation pass inside its temporary build VM while constructing the image — the same sops-nix activation-before-first-boot problem the installer and LXC tarball workflows have (see docs/auto-installer.md) applies here too, unmodified. Disko has a native mechanism for it:

sudo ./result \
  --pre-format-files host-keys/server_ssh_host_ed25519_key /etc/ssh/ssh_host_ed25519_key \
  --pre-format-files host-keys/server_ssh_host_ed25519_key.pub /etc/ssh/ssh_host_ed25519_key.pub \
  --build-memory 2048

Generate the key first with scripts/secrets/sync-host-keys.sh <hostname>, same as any other host — see docs/auto-installer.md for the full walkthrough (it registers the new key in .sops.yaml and re-encrypts the affected secrets/*.yaml files too, no manual editing needed).

Deploying to Proxmox

The image needs UEFI (OVMF), not Proxmox's default SeaBIOS — modules/boot/efi.nix uses systemd-boot, which only works with UEFI firmware. virtio-scsi is safe to use as the disk bus: hardware-configuration/vm/proxmox.nix already includes virtio_scsi in its initrd kernel modules.

1. Copy the image to the Proxmox host:

   scp server.raw root@<proxmox-host>:/var/lib/vz/import/
   

2. Create an empty VM shell (no disk yet) — replace <vmid> with a free ID and <storage> with your storage pool's name (pvesm status or Datacenter → Storage in the web UI):

   qm create <vmid> --name proxmox-server --memory 2048 --cores 2 \
     --net0 virtio,bridge=vmbr0 \
     --bios ovmf --machine q35 \
     --scsihw virtio-scsi-pci \
     --efidisk0 <storage>:1,efitype=4m,pre-enrolled-keys=0
   

   (--efidisk0 is required for UEFI — it's where OVMF persists boot-entry NVRAM; without it, systemd-boot's boot entry may not survive a reboot.)

3. Import the raw disk into storage:

   qm importdisk <vmid> /var/lib/vz/import/server.raw <storage>
   

   This prints the resulting disk identifier (e.g. vm-<vmid>-disk-1).

4. Attach it and set it as the boot disk:

   qm set <vmid> --scsi0 <storage>:vm-<vmid>-disk-1
   qm set <vmid> --boot order=scsi0
   

5. Boot it:

   qm start <vmid>
   

No install step — it boots straight into the already-activated system.

Why not nix build .#nixosConfigurations.<host>.config.system.build.vm?

That's a different, unrelated feature — system.build.vm (nixos-rebuild build-vm) produces an ephemeral QEMU script for locally testing a configuration, not a distributable disk image. It's not part of this workflow.