scripts/installer/auto-install.sh

@@ -1,17 +1,20 @@
 #!/usr/bin/env nix-shell
-#!nix-shell -i bash -p jq disko nixos-install-tools
+#!nix-shell -i bash -p jq disko nixos-install-tools zfs
 # shellcheck shell=bash
 # The only genuinely external tools this script calls directly: `jq`
-# (parsing the `nix eval` host list) and `disko`/`nixos-install` (the
+# (parsing the `nix eval` host list), `disko`/`nixos-install` (the
-# install itself). Everything disko shells out to internally
+# install itself), and `zpool` (exporting a ZFS root pool before reboot,
-# (parted/sgdisk/mkfs.*/zfs/...) is self-contained -- disko's own
+# see the comment above that call below). Everything disko shells out to
-# generated scripts hardcode absolute Nix store paths for those, they
+# internally (parted/sgdisk/mkfs.*/zfs/...) is self-contained -- disko's
+# own generated scripts hardcode absolute Nix store paths for those, they
 # don't rely on this script's PATH at all (confirmed by inspecting a
 # generated system.build.formatScript). The built installer image
-# (modules/installer/common.nix) already has all three in
+# (modules/installer/common.nix, plus the upstream
-# environment.systemPackages, so this nix-shell wrapper is a fast no-op
+# installation-cd-minimal.nix it imports via iso.nix) already has all
-# there; it's what makes the script also work standalone (e.g. run
+# four in environment.systemPackages, so this nix-shell wrapper is a
-# directly from a checkout on a stock ISO), where they aren't.
+# fast no-op there; it's what makes the script also work standalone
+# (e.g. run directly from a checkout on a stock ISO), where they aren't
+# guaranteed.
 set -eux
 
 set -euo pipefail
@@ -162,5 +165,24 @@ nixos-install \
     # /etc/ssh/ssh_host_ed25519_key. Nothing NixOS-managed ever cleans this
     # up on its own since it was written imperatively, not declaratively.
     rm -rf /root/host-keys
+
+    # disko's --mode ...,mount left any ZFS root pool imported (that's what
+    # let nixos-install write into /mnt). If we reboot with it still
+    # imported, it isn't just "not exported" -- it's stamped with *this*
+    # live installer environment's hostid, which almost never matches the
+    # target's own networking.hostId (see hosts/*/host.nix; the installer
+    # itself sets none). modules/services/zfs/enable-service.nix and
+    # modules/common/configuration.nix both set boot.zfs.forceImportRoot =
+    # false deliberately (the safe option per that setting's own docs), so
+    # the freshly-installed system's first real boot sees a pool "in use by
+    # another system" and refuses to import it without -f -- which is what
+    # makes boot stall waiting on the ZFS import. Exporting here (a no-op
+    # if the chosen host has no ZFS root, e.g. proxmox-*/linode-*) clears
+    # that in-use state so the next import, from any hostid, succeeds.
+    if [[ -n "$(zpool list -H -o name 2>/dev/null)" ]]; then
+        echo "Exporting ZFS pool(s) before reboot..."
+        zpool export -a
+    fi
+
     sleep 10
     reboot