hst: expose scheduling policy

This is primarily useful for poorly written music players for now.

Signed-off-by: Ophestra <cat@gensokyo.uk>

11 files changed, 91 insertions, 13 deletions

diff --git a/cmd/hakurei/command.go b/cmd/hakurei/command.go
index 26c48d78..242e1389 100644
--- a/cmd/hakurei/command.go
+++ b/cmd/hakurei/command.go
@@ -88,6 +88,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 			flagGroups   command.RepeatableFlag
 			flagHomeDir  string
 			flagUserName string
+			flagSched    string
 
 			flagPrivateRuntime, flagPrivateTmpdir bool
 
@@ -131,7 +132,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					log.Fatal(optionalErrorUnwrap(err))
 					return err
 				} else if progPath, err = check.NewAbs(p); err != nil {
-					log.Fatal(err.Error())
+					log.Fatal(err)
 					return err
 				}
 			}
@@ -150,7 +151,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				et |= hst.EPipeWire
 			}
 
-			config := &hst.Config{
+			config := hst.Config{
 				ID:          flagID,
 				Identity:    flagIdentity,
 				Groups:      flagGroups,
@@ -177,6 +178,10 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				},
 			}
 
+			if err := config.SchedPolicy.UnmarshalText([]byte(flagSched)); err != nil {
+				log.Fatal(err)
+			}
+
 			// bind GPU stuff
 			if et&(hst.EX11|hst.EWayland) != 0 {
 				config.Container.Filesystem = append(config.Container.Filesystem, hst.FilesystemConfigJSON{FilesystemConfig: &hst.FSBind{
@@ -214,7 +219,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					homeDir = passwd.HomeDir
 				}
 				if a, err := check.NewAbs(homeDir); err != nil {
-					log.Fatal(err.Error())
+					log.Fatal(err)
 					return err
 				} else {
 					config.Container.Home = a
@@ -234,11 +239,11 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					config.SessionBus = dbus.NewConfig(flagID, true, flagDBusMpris)
 				} else {
 					if f, err := os.Open(flagDBusConfigSession); err != nil {
-						log.Fatal(err.Error())
+						log.Fatal(err)
 					} else {
 						decodeJSON(log.Fatal, "load session bus proxy config", f, &config.SessionBus)
 						if err = f.Close(); err != nil {
-							log.Fatal(err.Error())
+							log.Fatal(err)
 						}
 					}
 				}
@@ -246,11 +251,11 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				// system bus proxy is optional
 				if flagDBusConfigSystem != "nil" {
 					if f, err := os.Open(flagDBusConfigSystem); err != nil {
-						log.Fatal(err.Error())
+						log.Fatal(err)
 					} else {
 						decodeJSON(log.Fatal, "load system bus proxy config", f, &config.SystemBus)
 						if err = f.Close(); err != nil {
-							log.Fatal(err.Error())
+							log.Fatal(err)
 						}
 					}
 				}
@@ -266,7 +271,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				}
 			}
 
-			outcome.Main(ctx, msg, config, -1)
+			outcome.Main(ctx, msg, &config, -1)
 			panic("unreachable")
 		}).
 			Flag(&flagDBusConfigSession, "dbus-config", command.StringFlag("builtin"),
@@ -287,6 +292,8 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				"Container home directory").
 			Flag(&flagUserName, "u", command.StringFlag("chronos"),
 				"Passwd user name within sandbox").
+			Flag(&flagSched, "sched", command.StringFlag(""),
+				"Scheduling policy to set for the container").
 			Flag(&flagPrivateRuntime, "private-runtime", command.BoolFlag(false),
 				"Do not share XDG_RUNTIME_DIR between containers under the same identity").
 			Flag(&flagPrivateTmpdir, "private-tmpdir", command.BoolFlag(false),
diff --git a/cmd/hakurei/command_test.go b/cmd/hakurei/command_test.go
index 4edf4a0b..8b5ae638 100644
--- a/cmd/hakurei/command_test.go
+++ b/cmd/hakurei/command_test.go
@@ -36,7 +36,7 @@ Commands:
 		},
 		{
 			"run", []string{"run", "-h"}, `
-Usage:	hakurei run [-h | --help] [--dbus-config <value>] [--dbus-system <value>] [--mpris] [--dbus-log] [--id <value>] [-a <int>] [-g <value>] [-d <value>] [-u <value>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pipewire] [--pulse] COMMAND [OPTIONS]
+Usage:	hakurei run [-h | --help] [--dbus-config <value>] [--dbus-system <value>] [--mpris] [--dbus-log] [--id <value>] [-a <int>] [-g <value>] [-d <value>] [-u <value>] [--sched <value>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pipewire] [--pulse] COMMAND [OPTIONS]
 
 Flags:
   -X	Enable direct connection to X11
@@ -66,6 +66,8 @@ Flags:
     	Do not share TMPDIR between containers under the same identity
   -pulse
     	Enable PulseAudio compatibility daemon
+  -sched string
+    	Scheduling policy to set for the container
   -u string
     	Passwd user name within sandbox (default "chronos")
   -wayland
diff --git a/hst/config.go b/hst/config.go
index 51ac62ec..5c020636 100644
--- a/hst/config.go
+++ b/hst/config.go
@@ -6,6 +6,7 @@ import (
 	"strings"
 
 	"hakurei.app/container/check"
+	"hakurei.app/container/std"
 )
 
 // Config configures an application container.
@@ -103,6 +104,10 @@ type Config struct {
 	// Init user namespace supplementary groups inherited by all container processes.
 	Groups []string `json:"groups"`
 
+	// Scheduling policy to set for the container. The zero value retains the
+	// current scheduling policy.
+	SchedPolicy std.SchedPolicy `json:"sched_policy,omitempty"`
+
 	// High level configuration applied to the underlying [container].
 	Container *ContainerConfig `json:"container"`
 }
@@ -116,6 +121,10 @@ var (
 	// [Config.Identity] value.
 	ErrIdentityBounds = errors.New("identity out of bounds")
 
+	// ErrSchedPolicyBounds is returned by [Config.Validate] for an out of bounds
+	// [Config.SchedPolicy] value.
+	ErrSchedPolicyBounds = errors.New("scheduling policy out of bounds")
+
 	// ErrEnviron is returned by [Config.Validate] if an environment variable
 	// name contains '=' or NUL.
 	ErrEnviron = errors.New("invalid environment variable name")
@@ -138,6 +147,13 @@ func (config *Config) Validate() error {
 			Msg: "identity " + strconv.Itoa(config.Identity) + " out of range"}
 	}
 
+	if config.SchedPolicy < 0 || config.SchedPolicy > std.SCHED_LAST {
+		return &AppError{Step: "validate configuration", Err: ErrSchedPolicyBounds,
+			Msg: "scheduling policy " +
+				strconv.Itoa(int(config.SchedPolicy)) +
+				" out of range"}
+	}
+
 	if err := config.SessionBus.CheckInterfaces("session"); err != nil {
 		return err
 	}
diff --git a/hst/config_test.go b/hst/config_test.go
index 776c21e9..ac50f127 100644
--- a/hst/config_test.go
+++ b/hst/config_test.go
@@ -22,6 +22,10 @@ func TestConfigValidate(t *testing.T) {
 			Msg: "identity -1 out of range"}},
 		{"identity upper", &hst.Config{Identity: 10000}, &hst.AppError{Step: "validate configuration", Err: hst.ErrIdentityBounds,
 			Msg: "identity 10000 out of range"}},
+		{"sched lower", &hst.Config{SchedPolicy: -1}, &hst.AppError{Step: "validate configuration", Err: hst.ErrSchedPolicyBounds,
+			Msg: "scheduling policy -1 out of range"}},
+		{"sched upper", &hst.Config{SchedPolicy: 0xcafe}, &hst.AppError{Step: "validate configuration", Err: hst.ErrSchedPolicyBounds,
+			Msg: "scheduling policy 51966 out of range"}},
 		{"dbus session", &hst.Config{SessionBus: &hst.BusConfig{See: []string{""}}},
 			&hst.BadInterfaceError{Interface: "", Segment: "session"}},
 		{"dbus system", &hst.Config{SystemBus: &hst.BusConfig{See: []string{""}}},
diff --git a/internal/outcome/outcome.go b/internal/outcome/outcome.go
index 42b50389..cb86b390 100644
--- a/internal/outcome/outcome.go
+++ b/internal/outcome/outcome.go
@@ -99,12 +99,17 @@ func newOutcomeState(k syscallDispatcher, msg message.Msg, id *hst.ID, config *h
 		Shim: &shimParams{
 			PrivPID: k.getpid(),
 			Verbose: msg.IsVerbose(),
+
+			SchedPolicy: config.SchedPolicy,
 		},
 
-		ID:        id,
-		Identity:  config.Identity,
-		UserID:    hsu.MustID(msg),
-		Paths:     env.CopyPathsFunc(k.fatalf, k.tempdir, func(key string) string { v, _ := k.lookupEnv(key); return v }),
+		ID:       id,
+		Identity: config.Identity,
+		UserID:   hsu.MustID(msg),
+		Paths: env.CopyPathsFunc(k.fatalf, k.tempdir, func(key string) string {
+			v, _ := k.lookupEnv(key)
+			return v
+		}),
 		Container: config.Container,
 	}
 
diff --git a/internal/outcome/shim.go b/internal/outcome/shim.go
index 8fc46f2a..53b37310 100644
--- a/internal/outcome/shim.go
+++ b/internal/outcome/shim.go
@@ -274,6 +274,7 @@ func shimEntrypoint(k syscallDispatcher) {
 	cancelContainer.Store(&stop)
 	sp := shimPrivate{k: k, id: state.id}
 	z := container.New(ctx, msg)
+	z.SchedPolicy = state.Shim.SchedPolicy
 	z.Params = *stateParams.params
 	z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
 
diff --git a/nixos.nix b/nixos.nix
index 800c81e3..926a1d2a 100644
--- a/nixos.nix
+++ b/nixos.nix
@@ -139,6 +139,7 @@ in
                     inherit (app) identity groups enablements;
                     inherit (dbusConfig) session_bus system_bus;
                     direct_wayland = app.insecureWayland;
+                    sched_policy = app.schedPolicy;
 
                     container = {
                       inherit (app)
diff --git a/options.nix b/options.nix
index 581ce181..ffce4dd0 100644
--- a/options.nix
+++ b/options.nix
@@ -98,6 +98,7 @@ in
               ints
               str
               bool
+              enum
               package
               anything
               submodule
@@ -237,6 +238,22 @@ in
               };
               hostAbstract = mkEnableOption "share abstract unix socket scope";
 
+              schedPolicy = mkOption {
+                type = nullOr (enum [
+                  "fifo"
+                  "rr"
+                  "batch"
+                  "idle"
+                  "deadline"
+                  "ext"
+                ]);
+                default = null;
+                description = ''
+                  Scheduling policy to set for the container.
+                  The zero value retains the current scheduling policy.
+                '';
+              };
+
               nix = mkEnableOption "nix daemon access";
               mapRealUid = mkEnableOption "mapping to priv-user uid";
               device = mkEnableOption "access to all devices";
diff --git a/test/configuration.nix b/test/configuration.nix
index 5724ec4f..62d8239d 100644
--- a/test/configuration.nix
+++ b/test/configuration.nix
@@ -28,6 +28,15 @@
   # Automatically login on tty1 as a normal user:
   services.getty.autologinUser = "alice";
 
+  security.pam.loginLimits = [
+    {
+      domain = "@users";
+      item = "rtprio";
+      type = "-";
+      value = 1;
+    }
+  ];
+
   environment = {
     systemPackages = with pkgs; [
       # For D-Bus tests:
diff --git a/test/interactive/configuration.nix b/test/interactive/configuration.nix
index 49c90343..9d1bddd8 100644
--- a/test/interactive/configuration.nix
+++ b/test/interactive/configuration.nix
@@ -23,6 +23,14 @@
   security = {
     sudo.wheelNeedsPassword = false;
     rtkit.enable = true;
+    pam.loginLimits = [
+      {
+        domain = "@users";
+        item = "rtprio";
+        type = "-";
+        value = 1;
+      }
+    ];
   };
 
   services = {
diff --git a/test/test.py b/test/test.py
index 485e475f..9071f563 100644
--- a/test/test.py
+++ b/test/test.py
@@ -206,6 +206,14 @@ machine.wait_until_fails("pgrep foot", timeout=5)
 machine.wait_for_file("/tmp/shim-cont-unexpected-pid")
 print(machine.succeed('grep "shim: got SIGCONT from unexpected process$" /tmp/shim-cont-unexpected-pid'))
 
+# Check setscheduler:
+sched_unset = int(machine.succeed("sudo -u alice -i hakurei -v run cat /proc/self/sched | grep '^policy' | tr -d ' ' | cut -d ':' -f 2"))
+if sched_unset != 0:
+    raise Exception(f"unexpected unset policy: {sched_unset}")
+sched_idle = int(machine.succeed("sudo -u alice -i hakurei -v run --sched=idle cat /proc/self/sched | grep '^policy' | tr -d ' ' | cut -d ':' -f 2"))
+if sched_idle != 5:
+    raise Exception(f"unexpected idle policy: {sched_idle}")
+
 # Start app (foot) with Wayland enablement:
 swaymsg("exec ne-foot")
 wait_for_window(f"u0_a{hakurei_identity(0)}@machine")