power optimizations

configuration.nix

@@ -20,6 +20,7 @@
     ./modules/no-rgb.nix
     ./modules/security.nix
     ./modules/ntfy-alerts.nix
+    ./modules/power.nix
 
     ./services/postgresql.nix
     ./services/jellyfin.nix
@@ -91,13 +92,6 @@
 
   services.kmscon.enable = true;
 
-  systemd.targets = {
-    sleep.enable = false;
-    suspend.enable = false;
-    hibernate.enable = false;
-    hybrid-sleep.enable = false;
-  };
-
   # Disable serial getty on ttyS0 to prevent dmesg warnings
   systemd.services."serial-getty@ttyS0".enable = false;
 
@@ -109,12 +103,6 @@
     enable = false;
   };
 
-  powerManagement = {
-    powertop.enable = true;
-    enable = true;
-    cpuFreqGovernor = "powersave";
-  };
-
   # https://github.com/NixOS/nixpkgs/issues/101459#issuecomment-758306434
   security.pam.loginLimits = [
     {

modules/power.nix

@@ -0,0 +1,93 @@
+{
+  lib,
+  pkgs,
+  ...
+}:
+{
+  powerManagement = {
+    enable = true;
+    powertop.enable = true;
+    cpuFreqGovernor = "powersave";
+  };
+
+  # Always-on server: disable all sleep targets.
+  systemd.targets = {
+    sleep.enable = false;
+    suspend.enable = false;
+    hibernate.enable = false;
+    hybrid-sleep.enable = false;
+  };
+
+  boot.kernelParams = [
+    # Disable NMI watchdog at boot. Eliminates periodic perf-counter interrupts
+    # across all cores (~1 W). Safe: apcupsd provides hardware hang detection
+    # via UPS, and softlockup watchdog remains active.
+    "nmi_watchdog=0"
+
+    # Route kernel work items to already-busy CPUs rather than waking idle ones.
+    # Reduces C-state exit frequency at the cost of slightly higher latency on
+    # work items -- irrelevant for a server whose latency-sensitive paths are
+    # all in userspace (caddy, jellyfin).
+    "workqueue.power_efficient=1"
+
+    # Force PCIe ASPM on even if the BIOS doesn't advertise support. ASRock
+    # B550M Pro4 BIOS defaults are conservative; the Zen 3 root complex and
+    # all downstream devices (NVMe, AHCI, Intel NIC) support L1 substates.
+    # powertop auto-tune sets the policy to powersupersave at runtime, but
+    # without `force` the kernel may refuse to enable ASPM at all if the BIOS
+    # opted out, making the policy write a no-op.
+    "pcie_aspm=force"
+  ];
+
+  boot.kernel.sysctl = {
+    # Belt-and-suspenders: also set via boot param, but sysctl ensures it
+    # stays off if anything re-enables it at runtime.
+    "kernel.nmi_watchdog" = 0;
+  };
+
+  # Server has no audio consumers. Power-gate the HDA codec at module load
+  # rather than waiting for powertop auto-tune to do it after boot.
+  boot.extraModprobeConfig = ''
+    options snd_hda_intel power_save=1 power_save_controller=Y
+  '';
+
+  # Apply sysfs power knobs that powertop --auto-tune cannot reach (hardened
+  # kernel blocks debugfs mount, so powertop silently skips ASPM policy and
+  # may only lower EPP to balance_power instead of power).
+  #
+  # AMD pstate EPP "power": deepest P-states, fastest core parking. Safe because:
+  # - xmrig runs at Nice=19 / CPUSchedulingPolicy=idle and tolerates latency
+  # - web services (caddy, jellyfin) are I/O-bound; the ~50 us extra C-state
+  #   exit latency is invisible behind network RTT
+  # - Minecraft server benefits from single-thread boost, which pstate still
+  #   provides on demand even in "power" mode (just with slightly slower ramp)
+  #
+  # ASPM powersupersave: deepest PCIe link power states (L1.1/L1.2). The
+  # pcie_aspm=force boot param enables ASPM, but the runtime policy defaults
+  # to "default" which only uses L0s. powersupersave adds L1 substates for
+  # all downstream devices (NVMe, AHCI, NIC).
+  systemd.services.power-tune = {
+    description = "Apply power-saving sysfs knobs (EPP, ASPM policy)";
+    after = [ "multi-user.target" ];
+    wantedBy = [ "multi-user.target" ];
+    serviceConfig = {
+      Type = "oneshot";
+      RemainAfterExit = true;
+      ExecStart = lib.getExe (
+        pkgs.writeShellApplication {
+          name = "power-tune";
+          text = ''
+            # AMD pstate energy performance preference
+            for epp in /sys/devices/system/cpu/cpu*/cpufreq/energy_performance_preference; do
+              [ -f "$epp" ] && echo power > "$epp"
+            done
+
+            # PCIe ASPM policy
+            aspm=/sys/module/pcie_aspm/parameters/policy
+            [ -f "$aspm" ] && echo powersupersave > "$aspm"
+          '';
+        }
+      );
+    };
+  };
+}