container: expose priority and SCHED_OTHER policy

The more explicit API removes the arbitrary limit preventing use of SCHED_OTHER (referred to as SCHED_NORMAL in the kernel). This change also exposes priority value to set.

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

cmd/mbf/main.go

@@ -87,7 +87,7 @@ func main() {
 		}
 
 		if flagIdle {
-			pkg.SchedPolicy = std.SCHED_IDLE
+			pkg.SetSchedIdle = true
 		}
 
 		return

container/container.go

@@ -38,9 +38,13 @@ type (
 	Container struct {
 		// Whether the container init should stay alive after its parent terminates.
 		AllowOrphan bool
-		// Scheduling policy to set via sched_setscheduler(2). The zero value
-		// skips this call. Supported policies are [SCHED_BATCH], [SCHED_IDLE].
+		// Whether to set SchedPolicy and SchedPriority via sched_setscheduler(2).
+		SetScheduler bool
+		// Scheduling policy to set via sched_setscheduler(2).
 		SchedPolicy std.SchedPolicy
+		// Scheduling priority to set via sched_setscheduler(2). The zero value
+		// implies the minimum value supported by the current SchedPolicy.
+		SchedPriority std.Int
 		// Cgroup fd, nil to disable.
 		Cgroup *int
 		// ExtraFiles passed through to initial process in the container, with
@@ -373,7 +377,15 @@ func (p *Container) Start() error {
 
 			// sched_setscheduler: thread-directed but acts on all processes
 			// created from the calling thread
-			if p.SchedPolicy > 0 && p.SchedPolicy <= std.SCHED_LAST {
+			if p.SetScheduler {
+				if p.SchedPolicy < 0 || p.SchedPolicy > std.SCHED_LAST {
+					return &StartError{
+						Fatal: false,
+						Step:  "set scheduling policy",
+						Err:   EINVAL,
+					}
+				}
+
 				var param schedParam
 				if priority, err := p.SchedPolicy.GetPriorityMin(); err != nil {
 					return &StartError{
@@ -382,10 +394,13 @@ func (p *Container) Start() error {
 						Err:   err,
 					}
 				} else {
-					param.priority = priority
+					param.priority = max(priority, p.SchedPriority)
 				}
 
-				p.msg.Verbosef("setting scheduling policy %s", p.SchedPolicy)
+				p.msg.Verbosef(
+					"setting scheduling policy %s priority %d",
+					p.SchedPolicy, param.priority,
+				)
 				if err := schedSetscheduler(
 					0, // calling thread
 					p.SchedPolicy,
@@ -393,7 +408,7 @@ func (p *Container) Start() error {
 				); err != nil {
 					return &StartError{
 						Fatal: true,
-						Step:  "enforce landlock ruleset",
+						Step:  "set scheduling policy",
 						Err:   err,
 					}
 				}

container/std/syscall.go

@@ -134,7 +134,7 @@ func (policy SchedPolicy) GetPriorityMax() (Int, error) {
 			0, 0,
 		)
 		schedPriority[policy][0] = Int(priority)
-		if schedPriority[policy][0] < 0 {
+		if errno != 0 {
 			schedPriorityErr[policy][0] = errno
 		}
 	})
@@ -151,10 +151,9 @@ func (policy SchedPolicy) GetPriorityMin() (Int, error) {
 			0, 0,
 		)
 		schedPriority[policy][1] = Int(priority)
-		if schedPriority[policy][1] < 0 {
+		if errno != 0 {
 			schedPriorityErr[policy][1] = errno
 		}
 	})
 	return schedPriority[policy][1], schedPriorityErr[policy][1]
-
 }

container/syscall.go

@@ -63,12 +63,12 @@ type schedParam struct {
 //
 // [very subtle to use correctly]: https://www.openwall.com/lists/musl/2016/03/01/4
 func schedSetscheduler(tid int, policy std.SchedPolicy, param *schedParam) error {
-	if r, _, errno := Syscall(
+	if _, _, errno := Syscall(
 		SYS_SCHED_SETSCHEDULER,
 		uintptr(tid),
 		uintptr(policy),
 		uintptr(unsafe.Pointer(param)),
-	); r < 0 {
+	); errno != 0 {
 		return errno
 	}
 	return nil

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.SetScheduler = state.Shim.SchedPolicy > 0
 	z.SchedPolicy = state.Shim.SchedPolicy
 	z.Params = *stateParams.params
 	z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr

internal/pkg/exec.go

@@ -39,8 +39,8 @@ type ExecPath struct {
 	W bool
 }
 
-// SchedPolicy is the [container] scheduling policy.
-var SchedPolicy std.SchedPolicy
+// SetSchedIdle is whether to set [std.SCHED_IDLE] scheduling priority.
+var SetSchedIdle bool
 
 // PromoteLayers returns artifacts with identical-by-content layers promoted to
 // the highest priority instance, as if mounted via [ExecPath].
@@ -413,7 +413,8 @@ func (a *execArtifact) cure(f *FContext, hostNet bool) (err error) {
 	z.ParentPerm = 0700
 	z.HostNet = hostNet
 	z.Hostname = "cure"
-	z.SchedPolicy = SchedPolicy
+	z.SetScheduler = SetSchedIdle
+	z.SchedPolicy = std.SCHED_IDLE
 	if z.HostNet {
 		z.Hostname = "cure-net"
 	}

test/test.py

@@ -213,6 +213,9 @@ if sched_unset != 0:
 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}")
+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"))
+if sched_rr != 2:
+    raise Exception(f"unexpected round-robin policy: {sched_idle}")
 
 # Start app (foot) with Wayland enablement:
 swaymsg("exec ne-foot")