Major fixes to systemd cgroup handling

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

Closes: #507
Approved by: baude

5 files changed, 45 insertions, 20 deletions

diff --git a/libpod/container.go b/libpod/container.go
index 40e670300..6a48363c9 100644
--- a/libpod/container.go
+++ b/libpod/container.go
@@ -6,7 +6,6 @@ import (
 	"path/filepath"
 	"time"
 
-	"github.com/containerd/cgroups"
 	"github.com/containernetworking/cni/pkg/types"
 	cnitypes "github.com/containernetworking/cni/pkg/types/current"
 	"github.com/containernetworking/plugins/pkg/ns"
@@ -854,9 +853,15 @@ func (c *Container) NamespacePath(ns LinuxNS) (string, error) {
 }
 
 // CGroupPath returns a cgroups "path" for a given container.
-func (c *Container) CGroupPath() cgroups.Path {
-	// TODO add support for systemd cgroup paths
-	return cgroups.StaticPath(filepath.Join(c.config.CgroupParent, fmt.Sprintf("libpod-conmon-%s", c.ID())))
+func (c *Container) CGroupPath() (string, error) {
+	switch c.runtime.config.CgroupManager {
+	case CgroupfsCgroupsManager:
+		return filepath.Join(c.config.CgroupParent, fmt.Sprintf("libpod-conmon-%s", c.ID())), nil
+	case SystemdCgroupsManager:
+		return filepath.Join(c.config.CgroupParent, createUnitName("libpod", c.ID())), nil
+	default:
+		return "", errors.Wrapf(ErrInvalidArg, "unsupported CGroup manager %s in use", c.runtime.config.CgroupManager)
+	}
 }
 
 // RootFsSize returns the root FS size of the container
diff --git a/libpod/container_internal.go b/libpod/container_internal.go
index 644598322..1098dc66c 100644
--- a/libpod/container_internal.go
+++ b/libpod/container_internal.go
@@ -7,6 +7,7 @@ import (
 	"io"
 	"io/ioutil"
 	"os"
+	"path"
 	"path/filepath"
 	"regexp"
 	"strings"
@@ -1128,12 +1129,19 @@ func (c *Container) generateSpec(ctx context.Context) (*spec.Spec, error) {
 		g.AddProcessEnv("container", "libpod")
 	}
 
-	cgroupPath, err := c.CGroupPath()("")
-	if err != nil {
-		return nil, errors.Wrapf(err, "error retrieving CGroup path for container %s", c.ID())
+	if c.runtime.config.CgroupManager == SystemdCgroupsManager {
+		// When runc is set to use Systemd as a cgroup manager, it
+		// expects cgroups to be passed as follows:
+		// slice:prefix:name
+		g.SetLinuxCgroupsPath(path.Base(c.config.CgroupParent) + ":" + "libpod" + ":" + c.ID())
+	} else {
+		cgroupPath, err := c.CGroupPath()
+		if err != nil {
+			return nil, err
+		}
+		logrus.Debugf("Setting CGroup path for container %s to %s", c.ID(), cgroupPath)
+		g.SetLinuxCgroupsPath(cgroupPath)
 	}
-	logrus.Debugf("Setting CGroup path for container %s to %s", c.ID(), cgroupPath)
-	g.SetLinuxCgroupsPath(cgroupPath)
 
 	return g.Spec(), nil
 }
diff --git a/libpod/container_top.go b/libpod/container_top.go
index 16030e040..a83303036 100644
--- a/libpod/container_top.go
+++ b/libpod/container_top.go
@@ -27,9 +27,9 @@ func (c *Container) GetContainerPids() ([]string, error) {
 // Gets the pids for a container without locking.  should only be called from a func where
 // locking has already been established.
 func (c *Container) getContainerPids() ([]string, error) {
-	cgroupPath, err := c.CGroupPath()("")
+	cgroupPath, err := c.CGroupPath()
 	if err != nil {
-		return nil, errors.Wrapf(err, "error getting cgroup path for container %s", c.ID())
+		return nil, err
 	}
 
 	taskFile := filepath.Join("/sys/fs/cgroup/pids", cgroupPath, "tasks")
diff --git a/libpod/oci.go b/libpod/oci.go
index a07aa8d86..e1df807cd 100644
--- a/libpod/oci.go
+++ b/libpod/oci.go
@@ -321,19 +321,26 @@ func (r *OCIRuntime) createOCIContainer(ctr *Container, cgroupParent string) (er
 
 	// Move conmon to specified cgroup
 	if r.cgroupManager == SystemdCgroupsManager {
-		logrus.Infof("Running conmon under slice %s and unitName %s", cgroupParent, createUnitName("libpod-conmon", ctr.ID()))
-		if err = utils.RunUnderSystemdScope(cmd.Process.Pid, cgroupParent, createUnitName("libpod-conmon", ctr.ID())); err != nil {
+		unitName := createUnitName("libpod", 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 {
 			logrus.Warnf("Failed to add conmon to systemd sandbox cgroup: %v", err)
 		}
 	} else {
-		control, err := cgroups.New(cgroups.V1, cgroups.StaticPath(filepath.Join(cgroupParent, "/libpod-conmon-"+ctr.ID())), &spec.LinuxResources{})
+		cgroupPath, err := ctr.CGroupPath()
 		if err != nil {
-			logrus.Warnf("Failed to add conmon to cgroupfs sandbox cgroup: %v", err)
+			logrus.Errorf("Failed to generate CGroup path for conmon: %v", err)
 		} else {
-			// we need to remove this defer and delete the cgroup once conmon exits
-			// maybe need a conmon monitor?
-			if err := control.Add(cgroups.Process{Pid: cmd.Process.Pid}); err != nil {
+			control, err := cgroups.New(cgroups.V1, cgroups.StaticPath(cgroupPath), &spec.LinuxResources{})
+			if err != nil {
 				logrus.Warnf("Failed to add conmon to cgroupfs sandbox cgroup: %v", err)
+			} else {
+				// we need to remove this defer and delete the cgroup once conmon exits
+				// maybe need a conmon monitor?
+				if err := control.Add(cgroups.Process{Pid: cmd.Process.Pid}); err != nil {
+					logrus.Warnf("Failed to add conmon to cgroupfs sandbox cgroup: %v", err)
+				}
 			}
 		}
 	}
diff --git a/libpod/stats.go b/libpod/stats.go
index 33e57572c..52bcc901d 100644
--- a/libpod/stats.go
+++ b/libpod/stats.go
@@ -43,9 +43,14 @@ func (c *Container) GetContainerStats(previousStats *ContainerStats) (*Container
 		return stats, nil
 	}
 
-	cgroup, err := cgroups.Load(cgroups.V1, c.CGroupPath())
+	cgroupPath, err := c.CGroupPath()
 	if err != nil {
-		return stats, errors.Wrapf(err, "unable to load cgroup at %+v", c.CGroupPath())
+		return nil, err
+	}
+
+	cgroup, err := cgroups.Load(cgroups.V1, cgroups.StaticPath(cgroupPath))
+	if err != nil {
+		return stats, errors.Wrapf(err, "unable to load cgroup at %s", cgroupPath)
 	}
 
 	cgroupStats, err := cgroup.Stat()