Merge pull request #169613 from helsinki-systems/feat/systemd-oomd

nixos/doc/manual/from_md/release-notes/rl-2211.section.xml

@@ -717,6 +717,21 @@
           Add udev rules for the Teensy family of microcontrollers.
         </para>
       </listitem>
+      <listitem>
+        <para>
+          systemd-oomd is enabled by default. Depending on which systemd
+          units have <literal>ManagedOOMSwap=kill</literal> or
+          <literal>ManagedOOMMemoryPressure=kill</literal>, systemd-oomd
+          will SIGKILL all the processes under the appropriate
+          descendant cgroups when the configured limits are exceeded.
+          NixOS does currently not configure cgroups with oomd by
+          default, this can be enabled using
+          <link xlink:href="options.html#opt-systemd.oomd.enableRootSlice">systemd.oomd.enableRootSlice</link>,
+          <link xlink:href="options.html#opt-systemd.oomd.enableSystemSlice">systemd.oomd.enableSystemSlice</link>,
+          and
+          <link xlink:href="options.html#opt-systemd.oomd.enableUserServices">systemd.oomd.enableUserServices</link>.
+        </para>
+      </listitem>
       <listitem>
         <para>
           The <literal>pass-secret-service</literal> package now

nixos/doc/manual/release-notes/rl-2211.section.md

@@ -235,6 +235,15 @@ Available as [services.patroni](options.html#opt-services.patroni.enable).
 
 - Add udev rules for the Teensy family of microcontrollers.
 
+- systemd-oomd is enabled by default. Depending on which systemd units have
+  `ManagedOOMSwap=kill` or `ManagedOOMMemoryPressure=kill`, systemd-oomd will
+  SIGKILL all the processes under the appropriate descendant cgroups when the
+  configured limits are exceeded. NixOS does currently not configure cgroups
+  with oomd by default, this can be enabled using
+  [systemd.oomd.enableRootSlice](options.html#opt-systemd.oomd.enableRootSlice),
+  [systemd.oomd.enableSystemSlice](options.html#opt-systemd.oomd.enableSystemSlice),
+  and [systemd.oomd.enableUserServices](options.html#opt-systemd.oomd.enableUserServices).
+
 - The `pass-secret-service` package now includes systemd units from upstream, so adding it to the NixOS `services.dbus.packages` option will make it start automatically as a systemd user service when an application tries to talk to the libsecret D-Bus API.
 
 - There is a new module for AMD SEV CPU functionality, which grants access to the hardware.

nixos/modules/module-list.nix

@@ -1236,6 +1236,7 @@
   ./system/boot/systemd/journald.nix
   ./system/boot/systemd/logind.nix
   ./system/boot/systemd/nspawn.nix
+  ./system/boot/systemd/oomd.nix
   ./system/boot/systemd/shutdown.nix
   ./system/boot/systemd/tmpfiles.nix
   ./system/boot/systemd/user.nix

nixos/modules/system/boot/systemd/oomd.nix

@@ -0,0 +1,57 @@
+{ config, lib, ... }: let
+
+  cfg = config.systemd.oomd;
+
+in {
+  options.systemd.oomd = {
+    enable = lib.mkEnableOption "the systemd-oomd OOM killer" // { default = true; };
+
+    # Fedora enables the first and third option by default. See the 10-oomd-* files here:
+    # https://src.fedoraproject.org/rpms/systemd/tree/acb90c49c42276b06375a66c73673ac351025597
+    enableRootSlice = lib.mkEnableOption "oomd on the root slice (-.slice)";
+    enableSystemSlice = lib.mkEnableOption "oomd on the system slice (system.slice)";
+    enableUserServices = lib.mkEnableOption "oomd on all user services (user@.service)";
+
+    extraConfig = lib.mkOption {
+      type = with lib.types; attrsOf (oneOf [ str int bool ]);
+      default = {};
+      example = lib.literalExpression ''{ DefaultMemoryPressureDurationSec = "20s"; }'';
+      description = ''
+        Extra config options for systemd-oomd. See man oomd.conf
+        for available options.
+      '';
+    };
+  };
+
+  config = lib.mkIf cfg.enable {
+    systemd.additionalUpstreamSystemUnits = [
+      "systemd-oomd.service"
+      "systemd-oomd.socket"
+    ];
+    systemd.services.systemd-oomd.wantedBy = [ "multi-user.target" ];
+
+    environment.etc."systemd/oomd.conf".text = lib.generators.toINI {} {
+      OOM = cfg.extraConfig;
+    };
+
+    systemd.oomd.extraConfig.DefaultMemoryPressureDurationSec = lib.mkDefault "20s"; # Fedora default
+
+    users.users.systemd-oom = {
+      description = "systemd-oomd service user";
+      group = "systemd-oom";
+      isSystemUser = true;
+    };
+    users.groups.systemd-oom = { };
+
+    systemd.slices."-".sliceConfig = lib.mkIf cfg.enableRootSlice {
+      ManagedOOMSwap = "kill";
+    };
+    systemd.slices."system".sliceConfig = lib.mkIf cfg.enableSystemSlice {
+      ManagedOOMSwap = "kill";
+    };
+    systemd.services."user@".serviceConfig = lib.mkIf cfg.enableUserServices {
+      ManagedOOMMemoryPressure = "kill";
+      ManagedOOMMemoryPressureLimit = "50%";
+    };
+  };
+}

nixos/tests/all-tests.nix

@@ -607,6 +607,7 @@ in {
   systemd-networkd-ipv6-prefix-delegation = handleTest ./systemd-networkd-ipv6-prefix-delegation.nix {};
   systemd-networkd-vrf = handleTest ./systemd-networkd-vrf.nix {};
   systemd-nspawn = handleTest ./systemd-nspawn.nix {};
+  systemd-oomd = handleTest ./systemd-oomd.nix {};
   systemd-shutdown = handleTest ./systemd-shutdown.nix {};
   systemd-timesyncd = handleTest ./systemd-timesyncd.nix {};
   systemd-misc = handleTest ./systemd-misc.nix {};

nixos/tests/systemd-oomd.nix

@@ -0,0 +1,37 @@
+import ./make-test-python.nix ({ pkgs, ... }:
+
+{
+  name = "systemd-oomd";
+
+  nodes.machine = { pkgs, ... }: {
+    systemd.oomd.extraConfig.DefaultMemoryPressureDurationSec = "1s"; # makes the test faster
+    # Kill cgroups when more than 1% pressure is encountered
+    systemd.slices."-".sliceConfig = {
+      ManagedOOMMemoryPressure = "kill";
+      ManagedOOMMemoryPressureLimit = "1%";
+    };
+    # A service to bring the system under memory pressure
+    systemd.services.testservice = {
+      serviceConfig.ExecStart = "${pkgs.coreutils}/bin/tail /dev/zero";
+    };
+    # Do not kill the backdoor
+    systemd.services.backdoor.serviceConfig.ManagedOOMMemoryPressure = "auto";
+
+    virtualisation.memorySize = 1024;
+  };
+
+  testScript = ''
+    # Start the system
+    machine.wait_for_unit("multi-user.target")
+    machine.succeed("oomctl")
+
+    # Bring the system into memory pressure
+    machine.succeed("echo 0 > /proc/sys/vm/panic_on_oom")  # NixOS tests kill the VM when the OOM killer is invoked - override this
+    machine.succeed("systemctl start testservice")
+
+    # Wait for oomd to kill something
+    # Matches these lines:
+    # systemd-oomd[508]: Killed /system.slice/systemd-udevd.service due to memory pressure for / being 3.26% > 1.00% for > 1s with reclaim activity
+    machine.wait_until_succeeds("journalctl -b | grep -q 'due to memory pressure for'")
+  '';
+})