Merge pull request #28566 from dubstack/dubstack-refactor-cgroup-manager

Automatic merge from submit-queue Add checks in Create and Update Cgroup methods This PR is connected to upstream issue for adding pod level cgroups in Kubernetes: #27204 Libcontainer currently doesen't support updates to parent devices cgroups. Until we get libcontainer to support skipping devices cgroup we will have that logic on the kubelet side. This PR includes: 1. Skip the devices cgroup when updating a cgroup. We only update the memory and cpu subsytems. 2. We explicitly pass all the cgroup paths that don't already exist to Apply() 3. Adds an AlreadyExists() method which is a utility function to check if all the subsystems of a cgroup already exist. On cgroupManager.Update() we only call Set() and cgroupManager.Create() we only call Apply() method @vishh PTAL
kubernetes · Jul 9, 2016 · 0f20ab6 · 0f20ab6
2 parents 516fba7 + dcfff45
commit 0f20ab6
Show file tree

Hide file tree

Showing 4 changed files with 106 additions and 67 deletions.
diff --git a/pkg/kubelet/cm/cgroup_manager_linux.go b/pkg/kubelet/cm/cgroup_manager_linux.go
@@ -18,7 +18,10 @@ package cm
 
 import (
 	"fmt"
+	"path"
 
+	libcontainercgroups "github.com/opencontainers/runc/libcontainer/cgroups"
+	cgroupfs "github.com/opencontainers/runc/libcontainer/cgroups/fs"
 	libcontainerconfigs "github.com/opencontainers/runc/libcontainer/configs"
 )
 
@@ -42,59 +45,133 @@ func NewCgroupManager(cs *cgroupSubsystems) CgroupManager {
 	}
 }
 
+// Exists checks if all subsystem cgroups already exist
+func (m *cgroupManagerImpl) Exists(name string) bool {
+	// Get map of all cgroup paths on the system for the particular cgroup
+	cgroupPaths := make(map[string]string, len(m.subsystems.mountPoints))
+	for key, val := range m.subsystems.mountPoints {
+		cgroupPaths[key] = path.Join(val, name)
+	}
+
+	// If even one cgroup doesn't exist we go on to create it
+	// @TODO(dubstack) We skip check for systemd until we update
+	// libcontainer in vendor
+	for key, path := range cgroupPaths {
+		if key != "systemd" && !libcontainercgroups.PathExists(path) {
+			return false
+		}
+	}
+	return true
+}
+
 // Destroy destroys the specified cgroup
 func (m *cgroupManagerImpl) Destroy(cgroupConfig *CgroupConfig) error {
 	//cgroup name
 	name := cgroupConfig.Name
 
-	// get the fscgroup Manager with the specified cgroup configuration
-	fsCgroupManager, err := getLibcontainerCgroupManager(cgroupConfig, m.subsystems)
+	// Get map of all cgroup paths on the system for the particular cgroup
+	cgroupPaths := make(map[string]string, len(m.subsystems.mountPoints))
+	for key, val := range m.subsystems.mountPoints {
+		cgroupPaths[key] = path.Join(val, name)
+	}
 
-	if err != nil {
-		return fmt.Errorf("Unable to destroy cgroup paths for cgroup %v : %v", name, err)
+	// Initialize libcontainer's cgroup config
+	libcontainerCgroupConfig := &libcontainerconfigs.Cgroup{
+		Name:   path.Base(name),
+		Parent: path.Dir(name),
+	}
+	fsCgroupManager := cgroupfs.Manager{
+		Cgroups: libcontainerCgroupConfig,
+		Paths:   cgroupPaths,
 	}
+
 	// Delete cgroups using libcontainers Managers Destroy() method
 	if err := fsCgroupManager.Destroy(); err != nil {
 		return fmt.Errorf("Unable to destroy cgroup paths for cgroup %v : %v", name, err)
 	}
 	return nil
 }
 
+type subsystem interface {
+	// Name returns the name of the subsystem.
+	Name() string
+	// Set the cgroup represented by cgroup.
+	Set(path string, cgroup *libcontainerconfigs.Cgroup) error
+}
+
+// Cgroup subsystems we currently support
+var supportedSubsystems []subsystem = []subsystem{
+	&cgroupfs.MemoryGroup{},
+	&cgroupfs.CpuGroup{},
+}
+
+// setSupportedSubsytems sets cgroup resource limits only on the supported
+// subsytems. ie. cpu and memory. We don't use libcontainer's cgroup/fs/Set()
+// method as it dosn't allow us to skip updates on the devices cgroup
+// Allowing or denying all devices by writing 'a' to devices.allow or devices.deny is
+// not possible once the device cgroups has children. Once the pod level cgroup are
+// created under the QOS level cgroup we cannot update the QOS level device cgroup.
+// We would like to skip setting any values on the device cgroup in this case
+// but this is not possible with libcontainers Set() method
+// See https://github.com/opencontainers/runc/issues/932
+func setSupportedSubsytems(cgroupConfig *libcontainerconfigs.Cgroup) error {
+	for _, sys := range supportedSubsystems {
+		if _, ok := cgroupConfig.Paths[sys.Name()]; !ok {
+			return fmt.Errorf("Failed to find subsytem mount for subsytem")
+		}
+		if err := sys.Set(cgroupConfig.Paths[sys.Name()], cgroupConfig); err != nil {
+			return fmt.Errorf("Failed to set config for supported subsystems : %v", err)
+		}
+	}
+	return nil
+}
+
 // Update updates the cgroup with the specified Cgroup Configuration
 func (m *cgroupManagerImpl) Update(cgroupConfig *CgroupConfig) error {
 	//cgroup name
 	name := cgroupConfig.Name
 
-	// get the fscgroup Manager with the specified cgroup configuration
-	fsCgroupManager, err := getLibcontainerCgroupManager(cgroupConfig, m.subsystems)
-	if err != nil {
-		return fmt.Errorf("Failed to update cgroup for %v : %v", name, err)
+	// Extract the cgroup resource parameters
+	resourceConfig := cgroupConfig.ResourceParameters
+	resources := &libcontainerconfigs.Resources{}
+	if resourceConfig.Memory != nil {
+		resources.Memory = *resourceConfig.Memory
+	}
+	if resourceConfig.CpuShares != nil {
+		resources.CpuShares = *resourceConfig.CpuShares
 	}
-	// get config object for passing to Set()
-	config := &libcontainerconfigs.Config{
-		Cgroups: fsCgroupManager.Cgroups,
+	if resourceConfig.CpuQuota != nil {
+		resources.CpuQuota = *resourceConfig.CpuQuota
 	}
 
-	// Update cgroup configuration using libcontainers Managers Set() method
-	if err := fsCgroupManager.Set(config); err != nil {
-		return fmt.Errorf("Failed to update cgroup for %v: %v", name, err)
+	// Initialize libcontainer's cgroup config
+	libcontainerCgroupConfig := &libcontainerconfigs.Cgroup{
+		Name:      path.Base(name),
+		Parent:    path.Dir(name),
+		Resources: resources,
+	}
+
+	if err := setSupportedSubsytems(libcontainerCgroupConfig); err != nil {
+		return fmt.Errorf("Failed to set supported cgroup subsystems for cgroup %v: %v", name, err)
 	}
 	return nil
 }
 
 // Create creates the specified cgroup
 func (m *cgroupManagerImpl) Create(cgroupConfig *CgroupConfig) error {
-	//cgroup name
+	// get cgroup name
 	name := cgroupConfig.Name
 
-	// get the fscgroup Manager with the specified cgroup configuration
-	fsCgroupManager, err := getLibcontainerCgroupManager(cgroupConfig, m.subsystems)
-	if err != nil {
-		return fmt.Errorf("Failed to create cgroup for %v : %v", name, err)
+	// Initialize libcontainer's cgroup config
+	libcontainerCgroupConfig := &libcontainerconfigs.Cgroup{
+		Name:      path.Base(name),
+		Parent:    path.Dir(name),
+		Resources: &libcontainerconfigs.Resources{},
 	}
-	// get config object for passing to libcontainer's Set() method
-	config := &libcontainerconfigs.Config{
-		Cgroups: fsCgroupManager.Cgroups,
+
+	// get the fscgroup Manager with the specified cgroup configuration
+	fsCgroupManager := &cgroupfs.Manager{
+		Cgroups: libcontainerCgroupConfig,
 	}
 	//Apply(0) is a hack to create the cgroup directories for each resource
 	// subsystem. The function [cgroups.Manager.apply()] applies cgroup
@@ -103,11 +180,7 @@ func (m *cgroupManagerImpl) Create(cgroupConfig *CgroupConfig) error {
 	// in the tasks file. We use the function to create all the required
 	// cgroup files but not attach any "real" pid to the cgroup.
 	if err := fsCgroupManager.Apply(0); err != nil {
-		return fmt.Errorf("Failed to create cgroup for %v: %v", name, err)
-	}
-	// Update cgroup configuration using libcontainers Managers Set() method
-	if err := fsCgroupManager.Set(config); err != nil {
-		return fmt.Errorf("Failed to create cgroup for %v: %v", name, err)
+		return fmt.Errorf("Failed to apply cgroup config for %v: %v", name, err)
 	}
 	return nil
 }
diff --git a/pkg/kubelet/cm/cgroup_manager_unsupported.go b/pkg/kubelet/cm/cgroup_manager_unsupported.go
@@ -29,6 +29,10 @@ func NewCgroupManager(_ interface{}) CgroupManager {
 	return &unsupportedCgroupManager{}
 }
 
+func (m *unsupportedCgroupManager) Exists(_ string) bool {
+	return false
+}
+
 func (m *unsupportedCgroupManager) Destroy(_ *CgroupConfig) error {
 	return nil
 }

diff --git a/pkg/kubelet/cm/helpers_linux.go b/pkg/kubelet/cm/helpers_linux.go
@@ -18,11 +18,8 @@ package cm
 
 import (
 	"fmt"
-	"path"
 
 	libcontainercgroups "github.com/opencontainers/runc/libcontainer/cgroups"
-	cgroupfs "github.com/opencontainers/runc/libcontainer/cgroups/fs"
-	libcontainerconfigs "github.com/opencontainers/runc/libcontainer/configs"
 )
 
 // cgroupSubsystems holds information about the mounted cgroup subsytems
@@ -59,40 +56,3 @@ func getCgroupSubsystems() (*cgroupSubsystems, error) {
 		mountPoints: mountPoints,
 	}, nil
 }
-
-// getLibcontainerCgroupManager returns libcontainer's cgroups manager
-// object with the specified cgroup configuration
-func getLibcontainerCgroupManager(cgroupConfig *CgroupConfig, subsystems *cgroupSubsystems) (*cgroupfs.Manager, error) {
-	// get cgroup name
-	name := cgroupConfig.Name
-
-	// Get map of all cgroup paths on the system for the particular cgroup
-	cgroupPaths := make(map[string]string, len(subsystems.mountPoints))
-	for key, val := range subsystems.mountPoints {
-		cgroupPaths[key] = path.Join(val, name)
-	}
-
-	// Extract the cgroup resource parameters
-	resourceConfig := cgroupConfig.ResourceParameters
-	resources := &libcontainerconfigs.Resources{}
-	resources.AllowAllDevices = true
-	if resourceConfig.Memory != nil {
-		resources.Memory = *resourceConfig.Memory
-	}
-	if resourceConfig.CpuShares != nil {
-		resources.CpuShares = *resourceConfig.CpuShares
-	}
-	if resourceConfig.CpuQuota != nil {
-		resources.CpuQuota = *resourceConfig.CpuQuota
-	}
-	// Initialize libcontainer's cgroup config
-	libcontainerCgroupConfig := &libcontainerconfigs.Cgroup{
-		Name:      path.Base(name),
-		Parent:    path.Dir(name),
-		Resources: resources,
-	}
-	return &cgroupfs.Manager{
-		Cgroups: libcontainerCgroupConfig,
-		Paths:   cgroupPaths,
-	}, nil
-}
diff --git a/pkg/kubelet/cm/types.go b/pkg/kubelet/cm/types.go
@@ -53,4 +53,6 @@ type CgroupManager interface {
 	Destroy(*CgroupConfig) error
 	// Update cgroup configuration.
 	Update(*CgroupConfig) error
+	// Exists checks if the cgroup already exists
+	Exists(string) bool
 }