diff --git a/cgroups/apply_raw.go b/cgroups/apply_raw.go
new file mode 100644
index 0000000..bce96f4
--- /dev/null
+++ b/cgroups/apply_raw.go
@@ -0,0 +1,189 @@
+package cgroups
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+)
+
+type rawCgroup struct {
+	root   string
+	cgroup string
+}
+
+func rawApply(c *Cgroup, 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/
+	//
+	// we can pick any subsystem to find the root
+
+	cgroupRoot, err := FindCgroupMountpoint("cpu")
+	if err != nil {
+		return nil, err
+	}
+	cgroupRoot = filepath.Dir(cgroupRoot)
+
+	if _, err := os.Stat(cgroupRoot); err != nil {
+		return nil, fmt.Errorf("cgroups fs not found")
+	}
+
+	cgroup := c.Name
+	if c.Parent != "" {
+		cgroup = filepath.Join(c.Parent, cgroup)
+	}
+
+	raw := &rawCgroup{
+		root:   cgroupRoot,
+		cgroup: cgroup,
+	}
+
+	if err := raw.setupDevices(c, pid); err != nil {
+		return nil, err
+	}
+	if err := raw.setupMemory(c, pid); err != nil {
+		return nil, err
+	}
+	if err := raw.setupCpu(c, pid); err != nil {
+		return nil, err
+	}
+	return raw, nil
+}
+
+func (raw *rawCgroup) path(subsystem string) (string, error) {
+	initPath, err := GetInitCgroupDir(subsystem)
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(raw.root, subsystem, initPath, raw.cgroup), nil
+}
+
+func (raw *rawCgroup) join(subsystem string, pid int) (string, error) {
+	path, err := raw.path(subsystem)
+	if err != nil {
+		return "", err
+	}
+	if err := os.MkdirAll(path, 0755); err != nil && !os.IsExist(err) {
+		return "", err
+	}
+	if err := writeFile(path, "tasks", strconv.Itoa(pid)); err != nil {
+		return "", err
+	}
+	return path, nil
+}
+
+func (raw *rawCgroup) setupDevices(c *Cgroup, pid int) (err error) {
+	if !c.DeviceAccess {
+		dir, err := raw.join("devices", pid)
+		if err != nil {
+			return err
+		}
+
+		defer func() {
+			if err != nil {
+				os.RemoveAll(dir)
+			}
+		}()
+
+		if err := writeFile(dir, "devices.deny", "a"); err != nil {
+			return err
+		}
+
+		allow := []string{
+			// /dev/null, zero, full
+			"c 1:3 rwm",
+			"c 1:5 rwm",
+			"c 1:7 rwm",
+
+			// consoles
+			"c 5:1 rwm",
+			"c 5:0 rwm",
+			"c 4:0 rwm",
+			"c 4:1 rwm",
+
+			// /dev/urandom,/dev/random
+			"c 1:9 rwm",
+			"c 1:8 rwm",
+
+			// /dev/pts/ - pts namespaces are "coming soon"
+			"c 136:* rwm",
+			"c 5:2 rwm",
+
+			// tuntap
+			"c 10:200 rwm",
+		}
+
+		for _, val := range allow {
+			if err := writeFile(dir, "devices.allow", val); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (raw *rawCgroup) setupMemory(c *Cgroup, pid int) (err error) {
+	if c.Memory != 0 || c.MemorySwap != 0 {
+		dir, err := raw.join("memory", pid)
+		if err != nil {
+			return err
+		}
+		defer func() {
+			if err != nil {
+				os.RemoveAll(dir)
+			}
+		}()
+
+		if c.Memory != 0 {
+			if err := writeFile(dir, "memory.limit_in_bytes", strconv.FormatInt(c.Memory, 10)); err != nil {
+				return err
+			}
+			if err := writeFile(dir, "memory.soft_limit_in_bytes", strconv.FormatInt(c.Memory, 10)); err != nil {
+				return err
+			}
+		}
+		// By default, MemorySwap is set to twice the size of RAM.
+		// If you want to omit MemorySwap, set it to `-1'.
+		if c.MemorySwap != -1 {
+			if err := writeFile(dir, "memory.memsw.limit_in_bytes", strconv.FormatInt(c.Memory*2, 10)); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (raw *rawCgroup) setupCpu(c *Cgroup, pid int) (err error) {
+	// We always want to join the cpu group, to allow fair cpu scheduling
+	// on a container basis
+	dir, err := raw.join("cpu", pid)
+	if err != nil {
+		return err
+	}
+	if c.CpuShares != 0 {
+		if err := writeFile(dir, "cpu.shares", strconv.FormatInt(c.CpuShares, 10)); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (raw *rawCgroup) Cleanup() error {
+	get := func(subsystem string) string {
+		path, _ := raw.path(subsystem)
+		return path
+	}
+
+	for _, path := range []string{
+		get("memory"),
+		get("devices"),
+		get("cpu"),
+	} {
+		if path != "" {
+			os.RemoveAll(path)
+		}
+	}
+	return nil
+}
diff --git a/cgroups/cgroups.go b/cgroups/cgroups.go
index b40e1a3..f35556f 100644
--- a/cgroups/cgroups.go
+++ b/cgroups/cgroups.go
@@ -8,7 +8,6 @@ import (
 	"io/ioutil"
 	"os"
 	"path/filepath"
-	"strconv"
 	"strings"
 )
 
@@ -22,6 +21,10 @@ type Cgroup struct {
 	CpuShares    int64 `json:"cpu_shares,omitempty"`    // CPU shares (relative weight vs. other containers)
 }
 
+type ActiveCgroup interface {
+	Cleanup() error
+}
+
 // https://www.kernel.org/doc/Documentation/cgroups/cgroups.txt
 func FindCgroupMountpoint(subsystem string) (string, error) {
 	mounts, err := mount.GetMounts()
@@ -62,48 +65,6 @@ func GetInitCgroupDir(subsystem string) (string, error) {
 	return parseCgroupFile(subsystem, f)
 }
 
-func (c *Cgroup) Path(root, subsystem string) (string, error) {
-	cgroup := c.Name
-	if c.Parent != "" {
-		cgroup = filepath.Join(c.Parent, cgroup)
-	}
-	initPath, err := GetInitCgroupDir(subsystem)
-	if err != nil {
-		return "", err
-	}
-	return filepath.Join(root, subsystem, initPath, cgroup), nil
-}
-
-func (c *Cgroup) Join(root, subsystem string, pid int) (string, error) {
-	path, err := c.Path(root, subsystem)
-	if err != nil {
-		return "", err
-	}
-	if err := os.MkdirAll(path, 0755); err != nil && !os.IsExist(err) {
-		return "", err
-	}
-	if err := writeFile(path, "tasks", strconv.Itoa(pid)); err != nil {
-		return "", err
-	}
-	return path, nil
-}
-
-func (c *Cgroup) Cleanup(root string) error {
-	get := func(subsystem string) string {
-		path, _ := c.Path(root, subsystem)
-		return path
-	}
-
-	for _, path := range []string{
-		get("memory"),
-		get("devices"),
-		get("cpu"),
-	} {
-		os.RemoveAll(path)
-	}
-	return nil
-}
-
 func parseCgroupFile(subsystem string, r io.Reader) (string, error) {
 	s := bufio.NewScanner(r)
 	for s.Scan() {
@@ -125,126 +86,6 @@ func writeFile(dir, file, data string) error {
 	return ioutil.WriteFile(filepath.Join(dir, file), []byte(data), 0700)
 }
 
-func (c *Cgroup) Apply(pid int) 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/
-	//
-	// we can pick any subsystem to find the root
-	cgroupRoot, err := FindCgroupMountpoint("cpu")
-	if err != nil {
-		return err
-	}
-	cgroupRoot = filepath.Dir(cgroupRoot)
-
-	if _, err := os.Stat(cgroupRoot); err != nil {
-		return fmt.Errorf("cgroups fs not found")
-	}
-	if err := c.setupDevices(cgroupRoot, pid); err != nil {
-		return err
-	}
-	if err := c.setupMemory(cgroupRoot, pid); err != nil {
-		return err
-	}
-	if err := c.setupCpu(cgroupRoot, pid); err != nil {
-		return err
-	}
-	return nil
-}
-
-func (c *Cgroup) setupDevices(cgroupRoot string, pid int) (err error) {
-	if !c.DeviceAccess {
-		dir, err := c.Join(cgroupRoot, "devices", pid)
-		if err != nil {
-			return err
-		}
-
-		defer func() {
-			if err != nil {
-				os.RemoveAll(dir)
-			}
-		}()
-
-		if err := writeFile(dir, "devices.deny", "a"); err != nil {
-			return err
-		}
-
-		allow := []string{
-			// /dev/null, zero, full
-			"c 1:3 rwm",
-			"c 1:5 rwm",
-			"c 1:7 rwm",
-
-			// consoles
-			"c 5:1 rwm",
-			"c 5:0 rwm",
-			"c 4:0 rwm",
-			"c 4:1 rwm",
-
-			// /dev/urandom,/dev/random
-			"c 1:9 rwm",
-			"c 1:8 rwm",
-
-			// /dev/pts/ - pts namespaces are "coming soon"
-			"c 136:* rwm",
-			"c 5:2 rwm",
-
-			// tuntap
-			"c 10:200 rwm",
-		}
-
-		for _, val := range allow {
-			if err := writeFile(dir, "devices.allow", val); err != nil {
-				return err
-			}
-		}
-	}
-	return nil
-}
-
-func (c *Cgroup) setupMemory(cgroupRoot string, pid int) (err error) {
-	if c.Memory != 0 || c.MemorySwap != 0 {
-		dir, err := c.Join(cgroupRoot, "memory", pid)
-		if err != nil {
-			return err
-		}
-		defer func() {
-			if err != nil {
-				os.RemoveAll(dir)
-			}
-		}()
-
-		if c.Memory != 0 {
-			if err := writeFile(dir, "memory.limit_in_bytes", strconv.FormatInt(c.Memory, 10)); err != nil {
-				return err
-			}
-			if err := writeFile(dir, "memory.soft_limit_in_bytes", strconv.FormatInt(c.Memory, 10)); err != nil {
-				return err
-			}
-		}
-		// By default, MemorySwap is set to twice the size of RAM.
-		// If you want to omit MemorySwap, set it to `-1'.
-		if c.MemorySwap != -1 {
-			if err := writeFile(dir, "memory.memsw.limit_in_bytes", strconv.FormatInt(c.Memory*2, 10)); err != nil {
-				return err
-			}
-		}
-	}
-	return nil
-}
-
-func (c *Cgroup) setupCpu(cgroupRoot string, pid int) (err error) {
-	// We always want to join the cpu group, to allow fair cpu scheduling
-	// on a container basis
-	dir, err := c.Join(cgroupRoot, "cpu", pid)
-	if err != nil {
-		return err
-	}
-	if c.CpuShares != 0 {
-		if err := writeFile(dir, "cpu.shares", strconv.FormatInt(c.CpuShares, 10)); err != nil {
-			return err
-		}
-	}
-	return nil
+func (c *Cgroup) Apply(pid int) (ActiveCgroup, error) {
+	return rawApply(c, pid)
 }
diff --git a/libcontainer/nsinit/exec.go b/libcontainer/nsinit/exec.go
index 61286cc..a44faaf 100644
--- a/libcontainer/nsinit/exec.go
+++ b/libcontainer/nsinit/exec.go
@@ -3,6 +3,7 @@
 package nsinit
 
 import (
+	"github.com/dotcloud/docker/pkg/cgroups"
 	"github.com/dotcloud/docker/pkg/libcontainer"
 	"github.com/dotcloud/docker/pkg/libcontainer/network"
 	"github.com/dotcloud/docker/pkg/system"
@@ -61,10 +62,15 @@ func (ns *linuxNs) Exec(container *libcontainer.Container, term Terminal, args [
 	// Do this before syncing with child so that no children
 	// can escape the cgroup
 	ns.logger.Println("setting cgroups")
-	if err := ns.SetupCgroups(container, command.Process.Pid); err != nil {
+	activeCgroup, err := ns.SetupCgroups(container, command.Process.Pid)
+	if err != nil {
 		command.Process.Kill()
 		return -1, err
 	}
+	if activeCgroup != nil {
+		defer activeCgroup.Cleanup()
+	}
+
 	ns.logger.Println("setting up network")
 	if err := ns.InitializeNetworking(container, command.Process.Pid, syncPipe); err != nil {
 		command.Process.Kill()
@@ -85,13 +91,11 @@ func (ns *linuxNs) Exec(container *libcontainer.Container, term Terminal, args [
 	return status, err
 }
 
-func (ns *linuxNs) SetupCgroups(container *libcontainer.Container, nspid int) error {
+func (ns *linuxNs) SetupCgroups(container *libcontainer.Container, nspid int) (cgroups.ActiveCgroup, error) {
 	if container.Cgroups != nil {
-		if err := container.Cgroups.Apply(nspid); err != nil {
-			return err
-		}
+		return container.Cgroups.Apply(nspid)
 	}
-	return nil
+	return nil, nil
 }
 
 func (ns *linuxNs) InitializeNetworking(container *libcontainer.Container, nspid int, pipe *SyncPipe) error {