Create pod CGroups when using the systemd cgroup driver

Signed-off-by: Matthew Heon <matthew.heon@gmail.com>

Closes: #1266
Approved by: baude

4 files changed, 109 insertions, 13 deletions

diff --git a/libpod/oci_linux.go b/libpod/oci_linux.go
index bdf859d7d..dacc92c62 100644
--- a/libpod/oci_linux.go
+++ b/libpod/oci_linux.go
@@ -24,8 +24,14 @@ func (r *OCIRuntime) moveConmonToCgroup(ctr *Container, cgroupParent string, cmd
 		if r.cgroupManager == SystemdCgroupsManager {
 			unitName := createUnitName("libpod-conmon", ctr.ID())
 
-			logrus.Infof("Running conmon under slice %s and unitName %s", cgroupParent, unitName)
-			if err := utils.RunUnderSystemdScope(cmd.Process.Pid, cgroupParent, unitName); err != nil {
+			realCgroupParent := cgroupParent
+			splitParent := strings.Split(cgroupParent, "/")
+			if strings.HasSuffix(cgroupParent, ".slice") && len(splitParent) > 1 {
+				realCgroupParent = splitParent[len(splitParent)-1]
+			}
+
+			logrus.Infof("Running conmon under slice %s and unitName %s", realCgroupParent, unitName)
+			if err := utils.RunUnderSystemdScope(cmd.Process.Pid, realCgroupParent, unitName); err != nil {
 				logrus.Warnf("Failed to add conmon to systemd sandbox cgroup: %v", err)
 			}
 		} else {
diff --git a/libpod/pod_internal.go b/libpod/pod_internal.go
index c8d8405bb..1ba4487ab 100644
--- a/libpod/pod_internal.go
+++ b/libpod/pod_internal.go
@@ -2,6 +2,7 @@ package libpod
 
 import (
 	"context"
+	"fmt"
 	"path/filepath"
 	"strings"
 	"time"
@@ -67,8 +68,11 @@ func (p *Pod) refresh() error {
 	if p.config.UsePodCgroup {
 		switch p.runtime.config.CgroupManager {
 		case SystemdCgroupsManager:
-			// NOOP for now, until proper systemd cgroup management
-			// is implemented
+			cgroupPath, err := systemdSliceFromPath(p.config.CgroupParent, fmt.Sprintf("libpod_pod_%s", p.ID()))
+			if err != nil {
+				logrus.Errorf("Error creating CGroup for pod %s: %v", p.ID(), err)
+			}
+			p.state.CgroupPath = cgroupPath
 		case CgroupfsCgroupsManager:
 			p.state.CgroupPath = filepath.Join(p.config.CgroupParent, p.ID())
 
diff --git a/libpod/runtime_pod_linux.go b/libpod/runtime_pod_linux.go
index 58affa45f..3592c2fee 100644
--- a/libpod/runtime_pod_linux.go
+++ b/libpod/runtime_pod_linux.go
@@ -4,6 +4,7 @@ package libpod
 
 import (
 	"context"
+	"fmt"
 	"path"
 	"path/filepath"
 	"strings"
@@ -70,12 +71,23 @@ func (r *Runtime) NewPod(options ...PodCreateOption) (*Pod, error) {
 		} else if len(pod.config.CgroupParent) < 6 || !strings.HasSuffix(path.Base(pod.config.CgroupParent), ".slice") {
 			return nil, errors.Wrapf(ErrInvalidArg, "did not receive systemd slice as cgroup parent when using systemd to manage cgroups")
 		}
-		// Creating CGroup path is currently a NOOP until proper systemd
-		// cgroup management is merged
+		// If we are set to use pod cgroups, set the cgroup parent that
+		// all containers in the pod will share
+		if pod.config.UsePodCgroup {
+			cgroupPath, err := systemdSliceFromPath(pod.config.CgroupParent, fmt.Sprintf("libpod_pod_%s", pod.ID()))
+			if err != nil {
+				return nil, errors.Wrapf(err, "unable to create pod cgroup for pod %s", pod.ID())
+			}
+			pod.state.CgroupPath = cgroupPath
+		}
 	default:
 		return nil, errors.Wrapf(ErrInvalidArg, "unsupported CGroup manager: %s - cannot validate cgroup parent", r.config.CgroupManager)
 	}
 
+	if pod.config.UsePodCgroup {
+		logrus.Debugf("Got pod cgroup as %s", pod.state.CgroupPath)
+	}
+
 	if err := r.state.AddPod(pod); err != nil {
 		return nil, errors.Wrapf(err, "error adding pod to state")
 	}
@@ -91,11 +103,20 @@ func (r *Runtime) removePod(ctx context.Context, p *Pod, removeCtrs, force bool)
 		return ErrRuntimeStopped
 	}
 
+	if !p.valid {
+		if ok, _ := r.state.HasPod(p.ID()); !ok {
+			// Pod was either already removed, or never existed to
+			// begin with
+			return nil
+		}
+	}
+
 	p.lock.Lock()
 	defer p.lock.Unlock()
 
-	if !p.valid {
-		return ErrPodRemoved
+	// Force a pod update
+	if err := p.updatePod(); err != nil {
+		return err
 	}
 
 	ctrs, err := r.state.PodContainers(p)
@@ -216,20 +237,27 @@ func (r *Runtime) removePod(ctx context.Context, p *Pod, removeCtrs, force bool)
 
 	// Remove pod cgroup, if present
 	if p.state.CgroupPath != "" {
+		logrus.Debugf("Removing pod cgroup %s", p.state.CgroupPath)
+
 		switch p.runtime.config.CgroupManager {
 		case SystemdCgroupsManager:
-			// NOOP for now, until proper systemd cgroup management
-			// is implemented
+			if err := deleteSystemdCgroup(p.state.CgroupPath); err != nil {
+				// The pod is already almost gone.
+				// No point in hard-failing if we fail
+				// this bit of cleanup.
+				logrus.Errorf("Error deleting pod %s cgroup %s: %v", p.ID(), p.state.CgroupPath, err)
+			}
 		case CgroupfsCgroupsManager:
 			// Delete the cgroupfs cgroup
-			logrus.Debugf("Removing pod cgroup %s", p.state.CgroupPath)
-
 			cgroup, err := cgroups.Load(cgroups.V1, cgroups.StaticPath(p.state.CgroupPath))
 			if err != nil && err != cgroups.ErrCgroupDeleted {
 				return err
 			} else if err == nil {
 				if err := cgroup.Delete(); err != nil {
-					return err
+					// The pod is already almost gone.
+					// No point in hard-failing if we fail
+					// this bit of cleanup.
+					logrus.Errorf("Error deleting pod %s cgroup %s: %v", p.ID(), p.state.CgroupPath, err)
 				}
 			}
 		default:
diff --git a/libpod/util.go b/libpod/util.go
index 13235059f..5b16b5755 100644
--- a/libpod/util.go
+++ b/libpod/util.go
@@ -9,10 +9,12 @@ import (
 	"strings"
 	"time"
 
+	"github.com/containerd/cgroups"
 	"github.com/containers/image/signature"
 	"github.com/containers/image/types"
 	spec "github.com/opencontainers/runtime-spec/specs-go"
 	"github.com/pkg/errors"
+	"github.com/sirupsen/logrus"
 )
 
 // Runtime API constants
@@ -123,6 +125,62 @@ func WaitForFile(path string, timeout time.Duration) error {
 	}
 }
 
+// systemdSliceFromPath makes a new systemd slice under the given parent with
+// the given name.
+// The parent must be a slice. The name must NOT include ".slice"
+func systemdSliceFromPath(parent, name string) (string, error) {
+	cgroupPath, err := assembleSystemdCgroupName(parent, name)
+	if err != nil {
+		return "", err
+	}
+
+	logrus.Debugf("Created cgroup path %s for parent %s and name %s", cgroupPath, parent, name)
+
+	if err := makeSystemdCgroup(cgroupPath); err != nil {
+		return "", errors.Wrapf(err, "error creating cgroup %s", cgroupPath)
+	}
+
+	logrus.Debugf("Created cgroup %s", cgroupPath)
+
+	return cgroupPath, nil
+}
+
+// makeSystemdCgroup creates a systemd CGroup at the given location.
+func makeSystemdCgroup(path string) error {
+	controller, err := cgroups.NewSystemd(SystemdDefaultCgroupParent)
+	if err != nil {
+		return err
+	}
+
+	return controller.Create(path, &spec.LinuxResources{})
+}
+
+// deleteSystemdCgroup deletes the systemd cgroup at the given location
+func deleteSystemdCgroup(path string) error {
+	controller, err := cgroups.NewSystemd(SystemdDefaultCgroupParent)
+	if err != nil {
+		return err
+	}
+
+	return controller.Delete(path)
+}
+
+// assembleSystemdCgroupName creates a systemd cgroup path given a base and
+// a new component to add.
+// The base MUST be systemd slice (end in .slice)
+func assembleSystemdCgroupName(baseSlice, newSlice string) (string, error) {
+	const sliceSuffix = ".slice"
+
+	if !strings.HasSuffix(baseSlice, sliceSuffix) {
+		return "", errors.Wrapf(ErrInvalidArg, "cannot assemble cgroup path with base %q - must end in .slice", baseSlice)
+	}
+
+	noSlice := strings.TrimSuffix(baseSlice, sliceSuffix)
+	final := fmt.Sprintf("%s/%s-%s%s", baseSlice, noSlice, newSlice, sliceSuffix)
+
+	return final, nil
+}
+
 type byDestination []spec.Mount
 
 func (m byDestination) Len() int {