hst: expose scheduling priority

This is useful when limits are configured to allow it.

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

8 files changed, 39 insertions, 11 deletions

diff --git a/cmd/hakurei/command.go b/cmd/hakurei/command.go
index 242e1389..bf638e23 100644
--- a/cmd/hakurei/command.go
+++ b/cmd/hakurei/command.go
@@ -16,6 +16,7 @@ import (
 	"hakurei.app/command"
 	"hakurei.app/container/check"
 	"hakurei.app/container/fhs"
+	"hakurei.app/container/std"
 	"hakurei.app/hst"
 	"hakurei.app/internal/dbus"
 	"hakurei.app/internal/env"
@@ -88,7 +89,9 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 			flagGroups   command.RepeatableFlag
 			flagHomeDir  string
 			flagUserName string
-			flagSched    string
+
+			flagSchedPolicy   string
+			flagSchedPriority int
 
 			flagPrivateRuntime, flagPrivateTmpdir bool
 
@@ -178,9 +181,12 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				},
 			}
 
-			if err := config.SchedPolicy.UnmarshalText([]byte(flagSched)); err != nil {
+			if err := config.SchedPolicy.UnmarshalText(
+				[]byte(flagSchedPolicy),
+			); err != nil {
 				log.Fatal(err)
 			}
+			config.SchedPriority = std.Int(flagSchedPriority)
 
 			// bind GPU stuff
 			if et&(hst.EX11|hst.EWayland) != 0 {
@@ -292,8 +298,10 @@ 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(""),
+			Flag(&flagSchedPolicy, "policy", command.StringFlag(""),
 				"Scheduling policy to set for the container").
+			Flag(&flagSchedPriority, "priority", command.IntFlag(0),
+				"Scheduling priority 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 8b5ae638..67c80697 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>] [--sched <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>] [--policy <value>] [--priority <int>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pipewire] [--pulse] COMMAND [OPTIONS]
 
 Flags:
   -X	Enable direct connection to X11
@@ -60,14 +60,16 @@ Flags:
     	Allow owning MPRIS D-Bus path, has no effect if custom config is available
   -pipewire
     	Enable connection to PipeWire via SecurityContext
+  -policy string
+    	Scheduling policy to set for the container
+  -priority int
+    	Scheduling priority to set for the container
   -private-runtime
     	Do not share XDG_RUNTIME_DIR between containers under the same identity
   -private-tmpdir
     	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 5c020636..a21be790 100644
--- a/hst/config.go
+++ b/hst/config.go
@@ -104,9 +104,15 @@ 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.
+	// Scheduling policy to set for the container.
+	//
+	// The zero value retains the current scheduling policy.
 	SchedPolicy std.SchedPolicy `json:"sched_policy,omitempty"`
+	// Scheduling priority to set for the container.
+	//
+	// The zero value implies the minimum priority of the current SchedPolicy.
+	// Has no effect if SchedPolicy is zero.
+	SchedPriority std.Int `json:"sched_priority,omitempty"`
 
 	// High level configuration applied to the underlying [container].
 	Container *ContainerConfig `json:"container"`
diff --git a/internal/outcome/outcome.go b/internal/outcome/outcome.go
index cb86b390..c72f688f 100644
--- a/internal/outcome/outcome.go
+++ b/internal/outcome/outcome.go
@@ -100,7 +100,8 @@ func newOutcomeState(k syscallDispatcher, msg message.Msg, id *hst.ID, config *h
 			PrivPID: k.getpid(),
 			Verbose: msg.IsVerbose(),
 
-			SchedPolicy: config.SchedPolicy,
+			SchedPolicy:   config.SchedPolicy,
+			SchedPriority: config.SchedPriority,
 		},
 
 		ID:       id,
diff --git a/internal/outcome/shim.go b/internal/outcome/shim.go
index 6d3e8d8b..a35ea1fe 100644
--- a/internal/outcome/shim.go
+++ b/internal/outcome/shim.go
@@ -75,6 +75,8 @@ type shimParams struct {
 
 	// Copied from [hst.Config].
 	SchedPolicy std.SchedPolicy
+	// Copied from [hst.Config].
+	SchedPriority std.Int
 
 	// Outcome setup ops, contains setup state. Populated by outcome.finalise.
 	Ops []outcomeOp
@@ -276,6 +278,7 @@ func shimEntrypoint(k syscallDispatcher) {
 	z := container.New(ctx, msg)
 	z.SetScheduler = state.Shim.SchedPolicy > 0
 	z.SchedPolicy = state.Shim.SchedPolicy
+	z.SchedPriority = state.Shim.SchedPriority
 	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 926a1d2a..15ec5287 100644
--- a/nixos.nix
+++ b/nixos.nix
@@ -140,6 +140,7 @@ in
                     inherit (dbusConfig) session_bus system_bus;
                     direct_wayland = app.insecureWayland;
                     sched_policy = app.schedPolicy;
+                    sched_priority = app.schedPriority;
 
                     container = {
                       inherit (app)
diff --git a/options.nix b/options.nix
index ffce4dd0..f624b6f5 100644
--- a/options.nix
+++ b/options.nix
@@ -253,6 +253,13 @@ in
                   The zero value retains the current scheduling policy.
                 '';
               };
+              schedPriority = mkOption {
+                type = nullOr (ints.between 1 99);
+                default = null;
+                description = ''
+                  Scheduling priority to set for the container.
+                '';
+              };
 
               nix = mkEnableOption "nix daemon access";
               mapRealUid = mkEnableOption "mapping to priv-user uid";
diff --git a/test/test.py b/test/test.py
index 57cf2084..571b71d9 100644
--- a/test/test.py
+++ b/test/test.py
@@ -210,10 +210,10 @@ print(machine.succeed('grep "shim: got SIGCONT from unexpected process$" /tmp/sh
 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"))
+sched_idle = int(machine.succeed("sudo -u alice -i hakurei -v run --policy=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}")
-sched_rr = int(machine.succeed("sudo -u alice -i hakurei -v run --sched=rr cat /proc/self/sched | grep '^policy' | tr -d ' ' | cut -d ':' -f 2"))
+sched_rr = int(machine.succeed("sudo -u alice -i hakurei -v run --policy=rr cat /proc/self/sched | grep '^policy' | tr -d ' ' | cut -d ':' -f 2"))
 if sched_rr != 2:
     raise Exception(f"unexpected round-robin policy: {sched_idle}")