Store cgroup manager on a per-container basis

When we create a container, we assign a cgroup parent based on
the current cgroup manager in use. This parent is only usable
with the cgroup manager the container is created with, so if the
default cgroup manager is later changed or overridden, the
container will not be able to start.

To solve this, store the cgroup manager that created the
container in container configuration, so we can guarantee a
container with a systemd cgroup parent will always be started
with systemd cgroups.

Unfortunately, this is very difficult to test in CI, due to the
fact that we hard-code cgroup manager on all invocations of
Podman in CI.

Fixes containers#7830

Signed-off-by: Matthew Heon <[email protected]>

libpod/container.go

@@ -888,9 +888,22 @@ func (c *Container) NamespacePath(linuxNS LinuxNS) (string, error) { //nolint:in
 	return fmt.Sprintf("/proc/%d/ns/%s", c.state.PID, linuxNS.String()), nil
 }
 
+// CgroupManager returns the cgroup manager used by the given container.
+func (c *Container) CgroupManager() string {
+	cgroupManager := c.config.CgroupManager
+	if cgroupManager == "" {
+		cgroupManager = c.runtime.config.Engine.CgroupManager
+	}
+	return cgroupManager
+}
+
 // CGroupPath returns a cgroups "path" for a given container.
 func (c *Container) CGroupPath() (string, error) {
+	cgroupManager := c.CgroupManager()
+
 	switch {
+	case c.config.NoCgroups || c.config.CgroupsMode == "disabled":
+		return "", errors.Wrapf(define.ErrNoCgroups, "this container is not creating cgroups")
 	case c.config.CgroupsMode == cgroupSplit:
 		if c.config.CgroupParent != "" {
 			return "", errors.Errorf("cannot specify cgroup-parent with cgroup-mode %q", cgroupSplit)
@@ -906,9 +919,9 @@ func (c *Container) CGroupPath() (string, error) {
 			return "", errors.Errorf("invalid cgroup for conmon %q", cg)
 		}
 		return strings.TrimSuffix(cg, "/supervisor") + "/container", nil
-	case c.runtime.config.Engine.CgroupManager == config.CgroupfsCgroupsManager:
+	case cgroupManager == config.CgroupfsCgroupsManager:
 		return filepath.Join(c.config.CgroupParent, fmt.Sprintf("libpod-%s", c.ID())), nil
-	case c.runtime.config.Engine.CgroupManager == config.SystemdCgroupsManager:
+	case cgroupManager == config.SystemdCgroupsManager:
 		if rootless.IsRootless() {
 			uid := rootless.GetRootlessUID()
 			parts := strings.SplitN(c.config.CgroupParent, "/", 2)
@@ -922,7 +935,7 @@ func (c *Container) CGroupPath() (string, error) {
 		}
 		return filepath.Join(c.config.CgroupParent, createUnitName("libpod", c.ID())), nil
 	default:
-		return "", errors.Wrapf(define.ErrInvalidArg, "unsupported CGroup manager %s in use", c.runtime.config.Engine.CgroupManager)
+		return "", errors.Wrapf(define.ErrInvalidArg, "unsupported CGroup manager %s in use", cgroupManager)
 	}
 }
 

libpod/container_config.go

@@ -275,13 +275,16 @@ type ContainerMiscConfig struct {
 	StopTimeout uint `json:"stopTimeout,omitempty"`
 	// Time container was created
 	CreatedTime time.Time `json:"createdTime"`
+	// CgroupManager is the cgroup manager used to create this container.
+	// If empty, the runtime default will be used.
+	CgroupManager string `json:"cgroupManager,omitempty"`
 	// NoCgroups indicates that the container will not create CGroups. It is
 	// incompatible with CgroupParent.  Deprecated in favor of CgroupsMode.
 	NoCgroups bool `json:"noCgroups,omitempty"`
 	// CgroupsMode indicates how the container will create cgroups
 	// (disabled, no-conmon, enabled).  It supersedes NoCgroups.
 	CgroupsMode string `json:"cgroupsMode,omitempty"`
-	// Cgroup parent of the container
+	// Cgroup parent of the container.
 	CgroupParent string `json:"cgroupParent"`
 	// LogPath log location
 	LogPath string `json:"logPath"`

libpod/container_inspect.go

@@ -729,7 +729,7 @@ func (c *Container) generateInspectContainerHostConfig(ctrSpec *spec.Spec, named
 	// CGroup parent
 	// Need to check if it's the default, and not print if so.
 	defaultCgroupParent := ""
-	switch c.runtime.config.Engine.CgroupManager {
+	switch c.CgroupManager() {
 	case config.CgroupfsCgroupsManager:
 		defaultCgroupParent = CgroupfsDefaultCgroupParent
 	case config.SystemdCgroupsManager:
@@ -738,6 +738,7 @@ func (c *Container) generateInspectContainerHostConfig(ctrSpec *spec.Spec, named
 	if c.config.CgroupParent != defaultCgroupParent {
 		hostConfig.CgroupParent = c.config.CgroupParent
 	}
+	hostConfig.CgroupManager = c.CgroupManager()
 
 	// PID namespace mode
 	pidMode := ""

libpod/container_internal_linux.go

@@ -1968,6 +1968,7 @@ func (c *Container) getOCICgroupPath() (string, error) {
 	if err != nil {
 		return "", err
 	}
+	cgroupManager := c.CgroupManager()
 	switch {
 	case (rootless.IsRootless() && !unified) || c.config.NoCgroups:
 		return "", nil
@@ -1980,22 +1981,22 @@ func (c *Container) getOCICgroupPath() (string, error) {
 			return "", err
 		}
 		return filepath.Join(selfCgroup, "container"), nil
-	case c.runtime.config.Engine.CgroupManager == config.SystemdCgroupsManager:
+	case cgroupManager == config.SystemdCgroupsManager:
 		// When the OCI runtime is set to use Systemd as a cgroup manager, it
 		// expects cgroups to be passed as follows:
 		// slice:prefix:name
 		systemdCgroups := fmt.Sprintf("%s:libpod:%s", path.Base(c.config.CgroupParent), c.ID())
 		logrus.Debugf("Setting CGroups for container %s to %s", c.ID(), systemdCgroups)
 		return systemdCgroups, nil
-	case c.runtime.config.Engine.CgroupManager == config.CgroupfsCgroupsManager:
+	case cgroupManager == config.CgroupfsCgroupsManager:
 		cgroupPath, err := c.CGroupPath()
 		if err != nil {
 			return "", err
 		}
 		logrus.Debugf("Setting CGroup path for container %s to %s", c.ID(), cgroupPath)
 		return cgroupPath, nil
 	default:
-		return "", errors.Wrapf(define.ErrInvalidArg, "invalid cgroup manager %s requested", c.runtime.config.Engine.CgroupManager)
+		return "", errors.Wrapf(define.ErrInvalidArg, "invalid cgroup manager %s requested", cgroupManager)
 	}
 }
 

libpod/define/container_inspect.go

@@ -236,6 +236,9 @@ type InspectContainerHostConfig struct {
 	// include a Mounts field in inspect.
 	// Format: <src>:<destination>[:<comma-separated options>]
 	Binds []string `json:"Binds"`
+	// CgroupManager is the cgroup manager used by the container.
+	// At present, allowed values are either "cgroupfs" or "systemd".
+	CgroupManager string `json:"CgroupManager,omitempty"`
 	// CgroupMode is the configuration of the container's cgroup namespace.
 	// Populated as follows:
 	// private - a cgroup namespace has been created

libpod/oci_conmon_linux.go

@@ -57,7 +57,6 @@ type ConmonOCIRuntime struct {
 	path              string
 	conmonPath        string
 	conmonEnv         []string
-	cgroupManager     string
 	tmpDir            string
 	exitsDir          string
 	socketsDir        string
@@ -102,7 +101,6 @@ func newConmonOCIRuntime(name string, paths []string, conmonPath string, runtime
 	runtime.runtimeFlags = runtimeFlags
 
 	runtime.conmonEnv = runtimeCfg.Engine.ConmonEnvVars
-	runtime.cgroupManager = runtimeCfg.Engine.CgroupManager
 	runtime.tmpDir = runtimeCfg.Engine.TmpDir
 	runtime.logSizeMax = runtimeCfg.Containers.LogSizeMax
 	runtime.noPivot = runtimeCfg.Engine.NoPivotRoot
@@ -149,10 +147,6 @@ func newConmonOCIRuntime(name string, paths []string, conmonPath string, runtime
 	runtime.exitsDir = filepath.Join(runtime.tmpDir, "exits")
 	runtime.socketsDir = filepath.Join(runtime.tmpDir, "socket")
 
-	if runtime.cgroupManager != config.CgroupfsCgroupsManager && runtime.cgroupManager != config.SystemdCgroupsManager {
-		return nil, errors.Wrapf(define.ErrInvalidArg, "invalid cgroup manager specified: %s", runtime.cgroupManager)
-	}
-
 	// Create the exit files and attach sockets directories
 	if err := os.MkdirAll(runtime.exitsDir, 0750); err != nil {
 		// The directory is allowed to exist
@@ -1325,7 +1319,7 @@ func (r *ConmonOCIRuntime) sharedConmonArgs(ctr *Container, cuuid, bundlePath, p
 		args = append(args, rFlags...)
 	}
 
-	if r.cgroupManager == config.SystemdCgroupsManager && !ctr.config.NoCgroups && ctr.config.CgroupsMode != cgroupSplit {
+	if ctr.CgroupManager() == config.SystemdCgroupsManager && !ctr.config.NoCgroups && ctr.config.CgroupsMode != cgroupSplit {
 		args = append(args, "-s")
 	}
 
@@ -1437,8 +1431,10 @@ func (r *ConmonOCIRuntime) moveConmonToCgroupAndSignal(ctr *Container, cmd *exec
 	}
 
 	if mustCreateCgroup {
+		// TODO: This should be a switch - we are not guaranteed that
+		// there are only 2 valid cgroup managers
 		cgroupParent := ctr.CgroupParent()
-		if r.cgroupManager == config.SystemdCgroupsManager {
+		if ctr.CgroupManager() == config.SystemdCgroupsManager {
 			unitName := createUnitName("libpod-conmon", ctr.ID())
 
 			realCgroupParent := cgroupParent

libpod/runtime_ctr.go

@@ -208,6 +208,7 @@ func (r *Runtime) setupContainer(ctx context.Context, ctr *Container) (_ *Contai
 	// Check CGroup parent sanity, and set it if it was not set.
 	// Only if we're actually configuring CGroups.
 	if !ctr.config.NoCgroups {
+		ctr.config.CgroupManager = r.config.Engine.CgroupManager
 		switch r.config.Engine.CgroupManager {
 		case config.CgroupfsCgroupsManager:
 			if ctr.config.CgroupParent == "" {