cgroups: Add systemd implementation of cgroups

This implements cgroup.Apply() using the systemd apis.
We create a transient unit called "docker-$id.scope" that contains
the container processes. We also have a way to set unit specific
properties, currently only defining the Slice to put the
scope in.

Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)

cgroups/apply_nosystemd.go

@@ -0,0 +1,15 @@
+// +build !linux
+
+package cgroups
+
+import (
+	"fmt"
+)
+
+func useSystemd() bool {
+	return false
+}
+
+func systemdApply(c *Cgroup, pid int) (ActiveCgroup, error) {
+	return nil, fmt.Errorf("Systemd not supported")
+}

cgroups/apply_systemd.go

@@ -0,0 +1,158 @@
+// +build linux
+
+package cgroups
+
+import (
+	"fmt"
+	systemd1 "github.com/coreos/go-systemd/dbus"
+	"github.com/dotcloud/docker/pkg/systemd"
+	"github.com/godbus/dbus"
+	"path/filepath"
+	"strings"
+	"sync"
+)
+
+type systemdCgroup struct {
+}
+
+var (
+	connLock              sync.Mutex
+	theConn               *systemd1.Conn
+	hasStartTransientUnit bool
+)
+
+func useSystemd() bool {
+	if !systemd.SdBooted() {
+		return false
+	}
+
+	connLock.Lock()
+	defer connLock.Unlock()
+
+	if theConn == nil {
+		var err error
+		theConn, err = systemd1.New()
+		if err != nil {
+			return false
+		}
+
+		// Assume we have StartTransientUnit
+		hasStartTransientUnit = true
+
+		// But if we get UnknownMethod error we don't
+		if _, err := theConn.StartTransientUnit("test.scope", "invalid"); err != nil {
+			if dbusError, ok := err.(dbus.Error); ok {
+				if dbusError.Name == "org.freedesktop.DBus.Error.UnknownMethod" {
+					hasStartTransientUnit = false
+				}
+			}
+		}
+	}
+
+	return hasStartTransientUnit
+}
+
+type DeviceAllow struct {
+	Node        string
+	Permissions string
+}
+
+func getIfaceForUnit(unitName string) string {
+	if strings.HasSuffix(unitName, ".scope") {
+		return "Scope"
+	}
+	if strings.HasSuffix(unitName, ".service") {
+		return "Service"
+	}
+	return "Unit"
+}
+
+func systemdApply(c *Cgroup, pid int) (ActiveCgroup, error) {
+	unitName := c.Parent + "-" + c.Name + ".scope"
+	slice := "system.slice"
+
+	var properties []systemd1.Property
+
+	for _, v := range c.UnitProperties {
+		switch v[0] {
+		case "Slice":
+			slice = v[1]
+		default:
+			return nil, fmt.Errorf("Unknown unit propery %s", v[0])
+		}
+	}
+
+	properties = append(properties,
+		systemd1.Property{"Slice", dbus.MakeVariant(slice)},
+		systemd1.Property{"Description", dbus.MakeVariant("docker container " + c.Name)},
+		systemd1.Property{"PIDs", dbus.MakeVariant([]uint32{uint32(pid)})})
+
+	if !c.DeviceAccess {
+		properties = append(properties,
+			systemd1.Property{"DevicePolicy", dbus.MakeVariant("strict")},
+			systemd1.Property{"DeviceAllow", dbus.MakeVariant([]DeviceAllow{
+				{"/dev/null", "rwm"},
+				{"/dev/zero", "rwm"},
+				{"/dev/full", "rwm"},
+				{"/dev/random", "rwm"},
+				{"/dev/urandom", "rwm"},
+				{"/dev/tty", "rwm"},
+				{"/dev/console", "rwm"},
+				{"/dev/tty0", "rwm"},
+				{"/dev/tty1", "rwm"},
+				{"/dev/pts/ptmx", "rwm"},
+				// There is no way to add /dev/pts/* here atm, so we hack this manually below
+				// /dev/pts/* (how to add this?)
+				// Same with tuntap, which doesn't exist as a node most of the time
+			})})
+	}
+
+	if c.Memory != 0 {
+		properties = append(properties,
+			systemd1.Property{"MemoryLimit", dbus.MakeVariant(uint64(c.Memory))})
+	}
+	// TODO: MemorySwap not available in systemd
+
+	if c.CpuShares != 0 {
+		properties = append(properties,
+			systemd1.Property{"CPUShares", dbus.MakeVariant(uint64(c.CpuShares))})
+	}
+
+	if _, err := theConn.StartTransientUnit(unitName, "replace", properties...); err != nil {
+		return nil, err
+	}
+
+	// To work around the lack of /dev/pts/* support above we need to manually add these
+	// so, ask systemd for the cgroup used
+	props, err := theConn.GetUnitTypeProperties(unitName, getIfaceForUnit(unitName))
+	if err != nil {
+		return nil, err
+	}
+
+	cgroup := props["ControlGroup"].(string)
+
+	if !c.DeviceAccess {
+		mountpoint, err := FindCgroupMountpoint("devices")
+		if err != nil {
+			return nil, err
+		}
+
+		path := filepath.Join(mountpoint, cgroup)
+
+		// /dev/pts/*
+		if err := writeFile(path, "devices.allow", "c 136:* rwm"); err != nil {
+			return nil, err
+		}
+		// tuntap
+		if err := writeFile(path, "devices.allow", "c 10:200 rwm"); err != nil {
+			return nil, err
+		}
+	}
+
+	return &systemdCgroup{}, nil
+}
+
+func (c *systemdCgroup) Cleanup() error {
+	// systemd cleans up, we don't need to do anything
+	return nil
+}

cgroups/cgroups.go

@@ -19,6 +19,8 @@ type Cgroup struct {
 	Memory       int64 `json:"memory,omitempty"`        // Memory limit (in bytes)
 	MemorySwap   int64 `json:"memory_swap,omitempty"`   // Total memory usage (memory + swap); set `-1' to disable swap
 	CpuShares    int64 `json:"cpu_shares,omitempty"`    // CPU shares (relative weight vs. other containers)
+
+	UnitProperties [][2]string `json:"unit_properties,omitempty"` // systemd unit properties
 }
 
 type ActiveCgroup interface {
@@ -87,5 +89,14 @@ func writeFile(dir, file, data string) error {
 }
 
 func (c *Cgroup) Apply(pid int) (ActiveCgroup, error) {
+	// We have two implementation of cgroups support, one is based on
+	// systemd and the dbus api, and one is based on raw cgroup fs operations
+	// following the pre-single-writer model docs at:
+	// http://www.freedesktop.org/wiki/Software/systemd/PaxControlGroups/
+
+	if useSystemd() {
+		return systemdApply(c, pid)
+	} else {
 		return rawApply(c, pid)
 	}
+}