Merge pull request 'Worktree ha file server test' (#81) from worktree-ha-file-server-test into main

Reviewed-on: #81
This commit is contained in:
2026-07-27 02:18:06 +00:00
16 changed files with 1178 additions and 1 deletions
+43
View File
@@ -0,0 +1,43 @@
# HA file server build type: DRBD + XFS + LIO iSCSI + NFS, managed by
# Corosync + Pacemaker. Both ha-server-1 and ha-server-2 use this type.
#
# NFS start/stop:
# services.nfs.server.enable = true configures /etc/exports, wires up
# rpcbind, and loads kernel modules — but nfs-server.service.wantedBy is
# force-cleared so systemd does NOT auto-start it at boot. Pacemaker's
# ha-group resource group (configured by scripts/ha/cluster-init.sh)
# starts and stops nfs-server as part of the failover sequence after the
# XFS mount and iSCSI target are brought up on the new Active node.
#
# Beszel agent:
# Enabled here via enable-agent.nix. The agent KEY (used to pair with
# the Beszel hub) is not set yet — add it to hosts/ha-server-{1,2}/host.nix
# under services.beszel.agent.environment.KEY once the hub accepts the
# new agents, following the pattern in hosts/server/host.nix.
{ lib, vars, ... }:
{
imports = [
../ha/pacemaker-stack.nix
../ha/iscsi-target.nix
../ha/cluster-config.nix
../beszel/enable-agent.nix
];
services.nfs.server = {
enable = true;
exports = ''
${vars.haStorageRoot}/${vars.nfsShares.dockerConfig.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.dockerVolumes.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.dockerDatabases.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.nextcloudData.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.raspiVolumes.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.proxmoxIsos.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.proxmoxLxcImages.subpath} ${vars.lanCidr}${vars.nfsShares.options}
${vars.haStorageRoot}/${vars.nfsShares.pxebootImages.subpath} ${vars.lanCidr}${vars.nfsShares.options}
'';
};
# Pacemaker controls nfs-server — prevent systemd from starting it at boot
# on both nodes (only the Active node should be serving NFS).
systemd.services.nfs-server.wantedBy = lib.mkForce [ ];
}
+106
View File
@@ -0,0 +1,106 @@
# Cluster-wide HA config shared by both ha-server nodes.
#
# Covers everything that is identical on both nodes and references cluster
# topology (node IPs, hostnames, DRBD resource). Per-node identity
# (hostname, static IP, stateVersion) lives in hosts/ha-server-{1,2}/host.nix.
#
# Corosync authkey:
# /etc/corosync/authkey (mode 0400) is managed by sops-nix below.
# Bootstrap: run scripts/ha/cluster-init.sh on node1 to generate the key,
# then encrypt it with: sops -e --input-type binary /etc/corosync/authkey > secrets/ha-corosync-authkey
# Both host keys must be registered via sync-host-keys.sh first so both nodes can decrypt it.
#
# DRBD fencing:
# Production setting is resource-only: DRBD waits for the STONITH fence
# agent to confirm the peer is dead before promoting to Primary. This
# requires a working fence_pve_ssh STONITH resource in Pacemaker
# (see scripts/ha/cluster-enable-stonith.sh). On a fresh cluster with
# no fence device yet, temporarily change to dont-care and run
# cluster-enable-stonith.sh once the fence key is deployed.
{ lib, vars, ... }:
{
services.drbd = {
enable = true;
config = ''
global {
usage-count yes;
}
common {
net {
protocol C;
ping-int 1;
verify-alg sha256;
after-sb-0pri discard-zero-changes;
after-sb-1pri discard-secondary;
}
disk {
fencing resource-only;
}
}
resource ha-data {
volume 0 {
device /dev/drbd0;
disk /dev/sdb;
meta-disk internal;
}
on ${vars.haServer1Host} {
address ${vars.haServer1Ip}:${toString vars.ports.haServerDrbd};
}
on ${vars.haServer2Host} {
address ${vars.haServer2Ip}:${toString vars.ports.haServerDrbd};
}
}
'';
};
# /etc/corosync/authkey — sops binary secret, identical on both nodes.
# Decryptable by both ha-server host keys (added by sync-host-keys.sh).
sops.secrets.corosync_authkey = {
sopsFile = ../../secrets/ha-corosync-authkey;
format = "binary";
path = "/etc/corosync/authkey";
mode = "0400";
restartUnits = [ "corosync.service" ];
};
# NixOS common config enables NetworkManager by default; HA cluster nodes
# need stable static IPs with predictable interface names — NM is not suitable.
networking.networkmanager.enable = lib.mkForce false;
# services.corosync.enable is set by modules/ha/pacemaker-stack.nix.
services.corosync = {
clusterName = "ha-cluster";
nodelist = [
{ nodeid = 1; name = vars.haServer1Host; ring_addrs = [ vars.haServer1Ip ]; }
{ nodeid = 2; name = vars.haServer2Host; ring_addrs = [ vars.haServer2Ip ]; }
];
};
networking.firewall = {
allowedTCPPorts = [
vars.ports.haServerIscsi
vars.ports.haServerPacemakerRemoted
vars.ports.haServerPcsd
vars.ports.haServerDrbd
vars.ports.nfsRpcbind
vars.ports.nfsd
vars.ports.nfsMountd
];
allowedUDPPorts = [
vars.ports.haServerCorosync1
vars.ports.haServerCorosync2
vars.ports.haServerCorosyncCrypto
vars.ports.nfsRpcbind
vars.ports.nfsd
vars.ports.nfsMountd
];
extraCommands = ''
iptables -A INPUT -s ${vars.haServer1Ip}/32 -j ACCEPT
iptables -A INPUT -s ${vars.haServer2Ip}/32 -j ACCEPT
'';
};
}
+99
View File
@@ -0,0 +1,99 @@
# LIO iSCSI target service (targetctl) for NixOS HA clusters.
#
# Provides the targetctl.service that saves/restores LIO configuration from
# /etc/target/saveconfig.json. Pacemaker manages this service via its
# systemd resource agent (class="systemd" type="targetctl").
#
# Why ExecStop is not simply "targetctl save":
# targetctl save writes the LIO config to JSON but does NOT remove the LIO
# target from the kernel's configfs. As a result, any fileio backing store
# that LIO has open (e.g. iscsi-lun.img on an XFS-over-DRBD filesystem)
# stays referenced in the kernel. The subsequent XFS umount from the
# Filesystem OCF resource then returns EBUSY and either hangs for the full
# op-stop timeout or fails outright, blocking the entire failover.
#
# The ExecStop script here additionally tears down the kernel LIO state
# via rtslib_fb after saving, so the backing-store file descriptor is
# released and umount succeeds immediately.
#
# Empty-config guard:
# The save step is skipped when no iSCSI targets are currently active.
# This prevents the secondary node (where LIO was never started) from
# overwriting a valid saveconfig.json with an empty one when Pacemaker
# stops the iscsi-target resource as part of a failover or cleanup.
{ pkgs, ... }:
let
python3 = pkgs.python3.withPackages (ps: [ ps.rtslib-fb ]);
targetctl = "${pkgs.targetcli-fb}/bin/targetctl";
targetctlStop = pkgs.writeScript "targetctl-stop" ''
#!${python3}/bin/python3
import subprocess, sys
import rtslib_fb
root = rtslib_fb.RTSRoot()
targets = list(root.targets)
if targets:
subprocess.run(
["${targetctl}", "save", "/etc/target/saveconfig.json"],
capture_output=True,
)
print(f"saved {len(targets)} iSCSI target(s)")
else:
print("no active LIO targets saveconfig.json unchanged")
for target in targets:
try:
for tpg in list(target.tpgs):
tpg.enable = False
target.delete()
except Exception as e:
print(f"warn (target): {e}", file=sys.stderr)
for so in list(root.storage_objects):
try:
so.delete()
except Exception as e:
print(f"warn (backstore): {e}", file=sys.stderr)
print("LIO kernel target cleared")
'';
in
{
boot.kernelModules = [
"target_core_mod"
"iscsi_target_mod"
"target_core_file"
"target_core_pscsi"
"target_core_user"
"configfs"
];
systemd = {
mounts = [{
where = "/sys/kernel/config";
what = "configfs";
type = "configfs";
wantedBy = [ "multi-user.target" ];
before = [ "targetctl.service" ];
}];
services.targetctl = {
description = "LIO iSCSI target config save/restore";
wantedBy = [ "multi-user.target" ];
after = [ "sys-kernel-config.mount" "network.target" ];
requires = [ "sys-kernel-config.mount" ];
serviceConfig = {
Type = "oneshot";
RemainAfterExit = true;
ExecStart = "${targetctl} restore /etc/target/saveconfig.json";
ExecStop = "${targetctlStop}";
};
unitConfig.ConditionFileNotEmpty = "/etc/target/saveconfig.json";
};
tmpfiles.rules = [
"d /etc/target 0750 root root -"
"f /etc/target/saveconfig.json 0640 root root -"
];
};
environment.systemPackages = [ pkgs.targetcli-fb ];
}
+94
View File
@@ -0,0 +1,94 @@
# Pacemaker + Corosync HA stack for NixOS with known-good workarounds.
#
# Issues fixed here (confirmed through live testing on NixOS 25.11):
#
# 1. StateDirectory ownership reset: systemd's StateDirectory=pacemaker
# creates /var/lib/pacemaker owned root:root. pacemaker-based (the CIB
# daemon) runs as the hacluster user and calls pcmk__daemon_can_write,
# which requires the CIB directory to be owned by hacluster or be
# group-writable by haclient. Workaround: remove StateDirectory and let
# ExecStartPre create every required subdirectory with correct ownership.
#
# 2. HA_SBIN_DIR wrong path: ocf-shellfuncs sets HA_SBIN_DIR to the Nix
# store path of the resource-agents derivation's /sbin, which doesn't
# exist. The DRBD OCF agent uses ${HA_SBIN_DIR}/crm_master, so it exits
# 127 without this override. Fix: export HA_SBIN_DIR=/run/current-system/sw/bin.
#
# 3. Broad PATH for OCF agents: the resource executor (pacemaker-execd) runs
# OCF agent scripts as children. NixOS provides no implicit PATH for
# system services; without an explicit PATH the agents can't find ip, ss,
# mount, umount, drbdadm, etc.
#
# 4. FUSER=true: the Filesystem OCF agent calls check_binary $FUSER (default:
# fuser from psmisc), which is not installed. Setting FUSER=true makes
# check_binary succeed (true is always in PATH) and the subsequent
# "$FUSER -km $mountpoint" becomes a no-op. Pair with force_unmount=false
# on each Filesystem resource unless you want lazy unmount behaviour.
{ lib, pkgs, ... }:
let
ocfBinPath = lib.concatStringsSep ":" [
"${pkgs.iproute2}/bin"
"${pkgs.iproute2}/sbin"
"${pkgs.iputils}/bin"
"${pkgs.util-linux}/bin"
"${pkgs.util-linux}/sbin"
"${pkgs.gawk}/bin"
"${pkgs.gnugrep}/bin"
"${pkgs.gnused}/bin"
"${pkgs.coreutils}/bin"
"${pkgs.bash}/bin"
"${pkgs.procps}/bin"
"${pkgs.xfsprogs}/bin"
"${pkgs.drbd}/bin"
"${pkgs.python3}/bin"
"/run/current-system/sw/bin"
"/run/current-system/sw/sbin"
"/usr/local/sbin"
"/usr/local/bin"
"/usr/sbin"
"/usr/bin"
"/sbin"
"/bin"
];
# Single pre-start script: schemas symlink + directory ownership.
# Runs before pacemakerd so pacemaker-based finds hacluster-owned dirs.
preStartCmd = "${pkgs.bash}/bin/bash -c '"
+ "ln -sfn ${pkgs.pacemaker}/share/pacemaker /var/lib/pacemaker/schemas; "
+ "for d in /var/lib/pacemaker /var/lib/pacemaker/cib /var/lib/pacemaker/cores "
+ "/var/lib/pacemaker/pengine /var/lib/pacemaker/blackbox "
+ "/var/lib/pacemaker/hostcache; do "
+ "mkdir -p \"\\$d\" && chown hacluster:pacemaker \"\\$d\" && chmod 2770 \"\\$d\"; "
+ "done'";
ocfEnv = {
PATH = lib.mkForce ocfBinPath;
OCF_ROOT = "${pkgs.ocf-resource-agents}/usr/lib/ocf";
HA_SBIN_DIR = "/run/current-system/sw/bin";
FUSER = "true";
};
in
{
users.groups.haclient = { };
services.corosync.enable = true;
services.pacemaker.enable = true;
systemd.services = {
pacemaker = {
serviceConfig = {
StateDirectory = lib.mkForce "";
ExecStartPre = lib.mkBefore [ preStartCmd ];
};
environment = ocfEnv;
};
pacemaker-execd.environment = ocfEnv;
};
environment.systemPackages = with pkgs; [
corosync
pacemaker
ocf-resource-agents
];
}