virtcontainers/pkg/cgroups: implement cgroup manager

cgroup manager is in charge to create and setup cgroups for
virtual containers, for example it adds /dev/kvm and
/dev/vhost-net to the list of cgroup devices in order to have
virtual containers working.

fixes #2438
fixes #2419

Signed-off-by: Julio Montes <julio.montes@intel.com>

virtcontainers/pkg/cgroups/manager.go

@@ -0,0 +1,321 @@
+// Copyright (c) 2020 Intel Corporation
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+
+package cgroups
+
+import (
+	"bufio"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+
+	"github.com/kata-containers/runtime/virtcontainers/pkg/rootless"
+	libcontcgroups "github.com/opencontainers/runc/libcontainer/cgroups"
+	libcontcgroupsfs "github.com/opencontainers/runc/libcontainer/cgroups/fs"
+	libcontcgroupssystemd "github.com/opencontainers/runc/libcontainer/cgroups/systemd"
+	"github.com/opencontainers/runc/libcontainer/configs"
+	"github.com/opencontainers/runc/libcontainer/specconv"
+	"github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/sirupsen/logrus"
+	"golang.org/x/sys/unix"
+)
+
+type Config struct {
+	// Cgroups specifies specific cgroup settings for the various subsystems that the container is
+	// placed into to limit the resources the container has available
+	// If nil, New() will create one.
+	Cgroups *configs.Cgroup
+
+	// CgroupPaths contains paths to all the cgroups setup for a container. Key is cgroup subsystem name
+	// with the value as the path.
+	CgroupPaths map[string]string
+
+	// Resources represents the runtime resource constraints
+	Resources specs.LinuxResources
+
+	// CgroupPath is the OCI spec cgroup path
+	CgroupPath string
+}
+
+type Manager struct {
+	sync.Mutex
+	mgr libcontcgroups.Manager
+}
+
+const (
+	// file in the cgroup that contains the pids
+	cgroupProcs = "cgroup.procs"
+)
+
+var (
+	// If set to true, expects cgroupsPath to be of form "slice:prefix:name", otherwise cgroups creation will fail
+	systemdCgroup *bool
+
+	cgroupsLogger = logrus.WithField("source", "virtcontainers/pkg/cgroups")
+)
+
+func EnableSystemdCgroup() {
+	systemd := true
+	systemdCgroup = &systemd
+}
+
+func UseSystemdCgroup() bool {
+	if systemdCgroup != nil {
+		return *systemdCgroup
+	}
+	return false
+}
+
+// returns the list of devices that a hypervisor may need
+func hypervisorDevices() []specs.LinuxDeviceCgroup {
+	wildcard := int64(-1)
+	devicemapperMajor := int64(253)
+
+	devices := []specs.LinuxDeviceCgroup{}
+
+	devices = append(devices,
+		// hypervisor needs access to all devicemapper devices,
+		// since they can be hotplugged in the VM.
+		specs.LinuxDeviceCgroup{
+			Allow:  true,
+			Type:   "b",
+			Major:  &devicemapperMajor,
+			Minor:  &wildcard,
+			Access: "rwm",
+		})
+
+	// Processes running in a device-cgroup are constrained, they have acccess
+	// only to the devices listed in the devices.list file.
+	// In order to run Virtual Machines and create virtqueues, hypervisors
+	// need access to certain character devices in the host, like kvm and vhost-net.
+	hypervisorDevices := []string{
+		"/dev/kvm",       // To run virtual machines
+		"/dev/vhost-net", // To create virtqueues
+	}
+
+	for _, device := range hypervisorDevices {
+		var st unix.Stat_t
+		linuxDevice := specs.LinuxDeviceCgroup{
+			Allow:  true,
+			Access: "rwm",
+		}
+
+		if err := unix.Stat(device, &st); err != nil {
+			cgroupsLogger.WithError(err).WithField("device", device).Warn("Could not get device information")
+			continue
+		}
+
+		switch st.Mode & unix.S_IFMT {
+		case unix.S_IFCHR:
+			linuxDevice.Type = "c"
+		case unix.S_IFBLK:
+			linuxDevice.Type = "b"
+		}
+
+		major := int64(unix.Major(st.Rdev))
+		minor := int64(unix.Minor(st.Rdev))
+		linuxDevice.Major = &major
+		linuxDevice.Minor = &minor
+
+		devices = append(devices, linuxDevice)
+	}
+
+	return devices
+}
+
+// New creates a new CgroupManager
+func New(config *Config) (*Manager, error) {
+	var err error
+	useSystemdCgroup := UseSystemdCgroup()
+
+	devices := []specs.LinuxDeviceCgroup{}
+	copy(devices, config.Resources.Devices)
+	devices = append(devices, hypervisorDevices()...)
+	// Do not modify original devices
+	config.Resources.Devices = devices
+
+	newSpec := specs.Spec{
+		Linux: &specs.Linux{
+			Resources: &config.Resources,
+		},
+	}
+
+	rootless := rootless.IsRootless()
+
+	cgroups := config.Cgroups
+	cgroupPaths := config.CgroupPaths
+
+	// Create a new cgroup if the current one is nil
+	// this cgroups must be saved later
+	if cgroups == nil {
+		if config.CgroupPath == "" && !rootless {
+			cgroupsLogger.Warn("cgroups have not been created and cgroup path is empty")
+		}
+
+		newSpec.Linux.CgroupsPath, err = ValidCgroupPath(config.CgroupPath, useSystemdCgroup)
+		if err != nil {
+			return nil, fmt.Errorf("Invalid cgroup path: %v", err)
+		}
+
+		if cgroups, err = specconv.CreateCgroupConfig(&specconv.CreateOpts{
+			// cgroup name is taken from spec
+			CgroupName:       "",
+			UseSystemdCgroup: useSystemdCgroup,
+			Spec:             &newSpec,
+			RootlessCgroups:  rootless,
+		}); err != nil {
+			return nil, fmt.Errorf("Could not create cgroup config: %v", err)
+		}
+	}
+
+	// Set cgroupPaths to nil when the map is empty, it can and will be
+	// populated by `Manager.Apply()` when the runtime or any other process
+	// is moved to the cgroup.
+	if len(cgroupPaths) == 0 {
+		cgroupPaths = nil
+	}
+
+	if useSystemdCgroup {
+		systemdCgroupFunc, err := libcontcgroupssystemd.NewSystemdCgroupsManager()
+		if err != nil {
+			return nil, fmt.Errorf("Could not create systemd cgroup manager: %v", err)
+		}
+		libcontcgroupssystemd.UseSystemd()
+		return &Manager{
+			mgr: systemdCgroupFunc(cgroups, cgroupPaths),
+		}, nil
+	}
+
+	return &Manager{
+		mgr: &libcontcgroupsfs.Manager{
+			Cgroups:  cgroups,
+			Rootless: rootless,
+			Paths:    cgroupPaths,
+		},
+	}, nil
+}
+
+// read all the pids in cgroupPath
+func readPids(cgroupPath string) ([]int, error) {
+	pids := []int{}
+	f, err := os.Open(filepath.Join(cgroupPath, cgroupProcs))
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	buf := bufio.NewScanner(f)
+
+	for buf.Scan() {
+		if t := buf.Text(); t != "" {
+			pid, err := strconv.Atoi(t)
+			if err != nil {
+				return nil, err
+			}
+			pids = append(pids, pid)
+		}
+	}
+	return pids, nil
+}
+
+// write the pids into cgroup.procs
+func writePids(pids []int, cgroupPath string) error {
+	cgroupProcsPath := filepath.Join(cgroupPath, cgroupProcs)
+	for _, pid := range pids {
+		if err := ioutil.WriteFile(cgroupProcsPath,
+			[]byte(strconv.Itoa(pid)),
+			os.FileMode(0),
+		); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (m *Manager) logger() *logrus.Entry {
+	return cgroupsLogger.WithField("source", "cgroup-manager")
+}
+
+// move all the processes in the current cgroup to the parent
+func (m *Manager) moveToParent() error {
+	m.Lock()
+	defer m.Unlock()
+	for _, cgroupPath := range m.mgr.GetPaths() {
+		pids, err := readPids(cgroupPath)
+		if err != nil {
+			return err
+		}
+
+		if len(pids) == 0 {
+			// no pids in this cgroup
+			continue
+		}
+
+		cgroupParentPath := filepath.Dir(filepath.Clean(cgroupPath))
+		if err = writePids(pids, cgroupParentPath); err != nil {
+			if !strings.Contains(err.Error(), "no such process") {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// Add pid to cgroups
+func (m *Manager) Add(pid int) error {
+	if rootless.IsRootless() {
+		m.logger().Debug("Unable to setup add pids to cgroup: running rootless")
+		return nil
+	}
+
+	m.Lock()
+	defer m.Unlock()
+	return m.mgr.Apply(pid)
+}
+
+// Apply constraints
+func (m *Manager) Apply() error {
+	if rootless.IsRootless() {
+		m.logger().Debug("Unable to apply constraints: running rootless")
+		return nil
+	}
+
+	cgroups, err := m.GetCgroups()
+	if err != nil {
+		return err
+	}
+
+	m.Lock()
+	defer m.Unlock()
+	return m.mgr.Set(&configs.Config{
+		Cgroups: cgroups,
+	})
+}
+
+func (m *Manager) GetCgroups() (*configs.Cgroup, error) {
+	m.Lock()
+	defer m.Unlock()
+	return m.mgr.GetCgroups()
+}
+
+func (m *Manager) GetPaths() map[string]string {
+	m.Lock()
+	defer m.Unlock()
+	return m.mgr.GetPaths()
+}
+
+func (m *Manager) Destroy() error {
+	// cgroup can't be destroyed if it contains running processes
+	if err := m.moveToParent(); err != nil {
+		return fmt.Errorf("Could not move processes into parent cgroup: %v", err)
+	}
+
+	m.Lock()
+	defer m.Unlock()
+	return m.mgr.Destroy()
+}

virtcontainers/pkg/cgroups/manager_test.go

@@ -0,0 +1,55 @@
+// Copyright (c) 2020 Intel Corporation
+//
+// SPDX-License-Identifier: Apache-2.0
+//
+
+package cgroups
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestEnableSystemdCgroup(t *testing.T) {
+	assert := assert.New(t)
+
+	orgSystemdCgroup := systemdCgroup
+	defer func() {
+		systemdCgroup = orgSystemdCgroup
+	}()
+
+	useSystemdCgroup := UseSystemdCgroup()
+	if systemdCgroup != nil {
+		assert.Equal(*systemdCgroup, useSystemdCgroup)
+	} else {
+		assert.False(useSystemdCgroup)
+	}
+
+	EnableSystemdCgroup()
+	assert.True(UseSystemdCgroup())
+}
+
+func TestNew(t *testing.T) {
+	assert := assert.New(t)
+	useSystemdCgroup := false
+	orgSystemdCgroup := systemdCgroup
+	defer func() {
+		systemdCgroup = orgSystemdCgroup
+	}()
+	systemdCgroup = &useSystemdCgroup
+
+	c := &Config{
+		Cgroups:    nil,
+		CgroupPath: "",
+	}
+
+	mgr, err := New(c)
+	assert.NoError(err)
+	assert.NotNil(mgr.mgr)
+
+	useSystemdCgroup = true
+	mgr, err = New(c)
+	assert.Error(err)
+	assert.Nil(mgr)
+}

virtcontainers/sandbox.go

@@ -34,6 +34,7 @@ import (
 	"github.com/kata-containers/runtime/virtcontainers/persist"
 	persistapi "github.com/kata-containers/runtime/virtcontainers/persist/api"
 	"github.com/kata-containers/runtime/virtcontainers/pkg/annotations"
+	vccgroups "github.com/kata-containers/runtime/virtcontainers/pkg/cgroups"
 	"github.com/kata-containers/runtime/virtcontainers/pkg/compatoci"
 	"github.com/kata-containers/runtime/virtcontainers/pkg/rootless"
 	vcTypes "github.com/kata-containers/runtime/virtcontainers/pkg/types"
@@ -214,6 +215,8 @@ type Sandbox struct {
 	seccompSupported  bool
 	disableVMShutdown bool
 
+	cgroupMgr *vccgroups.Manager
+
 	ctx context.Context
 }
 
@@ -597,6 +600,10 @@ func newSandbox(ctx context.Context, sandboxConfig SandboxConfig, factory Factor
 		}
 	}
 
+	if err := s.createCgroupManager(); err != nil {
+		return nil, err
+	}
+
 	agentConfig, err := newAgentConfig(sandboxConfig.AgentType, sandboxConfig.AgentConfig)
 	if err != nil {
 		return nil, err
@@ -609,6 +616,46 @@ func newSandbox(ctx context.Context, sandboxConfig SandboxConfig, factory Factor
 	return s, nil
 }
 
+func (s *Sandbox) createCgroupManager() error {
+	var err error
+	cgroupPath := ""
+
+	// Do not change current cgroup configuration.
+	// Create a spec without constraints
+	resources := specs.LinuxResources{}
+
+	if s.config == nil {
+		return fmt.Errorf("Could not create cgroup manager: empty sandbox configuration")
+	}
+
+	spec := s.GetPatchedOCISpec()
+	if spec != nil {
+		cgroupPath = spec.Linux.CgroupsPath
+
+		// kata should rely on the cgroup created and configured by
+		// container engine *only* if actual container was
+		// marked *explicitly* as sandbox through annotations.
+		if !s.config.HasCRIContainerType {
+			resources = *spec.Linux.Resources
+		}
+	}
+
+	// Create the cgroup manager, this way it can be used later
+	// to create or detroy cgroups
+	if s.cgroupMgr, err = vccgroups.New(
+		&vccgroups.Config{
+			Cgroups:     s.config.Cgroups,
+			CgroupPaths: s.state.CgroupPaths,
+			Resources:   resources,
+			CgroupPath:  cgroupPath,
+		},
+	); err != nil {
+		return err
+	}
+
+	return nil
+}
+
 // storeSandbox stores a sandbox config.
 func (s *Sandbox) storeSandbox() error {
 	span, _ := s.trace("storeSandbox")
@@ -1855,15 +1902,13 @@ func (s *Sandbox) cgroupsDelete() error {
 	var cgroupSubsystems cgroups.Hierarchy
 
 	if s.config.SandboxCgroupOnly {
-		cgroupSubsystems = cgroups.V1
-		path = s.state.CgroupPath
-		s.Logger().WithField("path", path).Debug("Deleting sandbox cgroups (all subsystems)")
-	} else {
-		cgroupSubsystems = V1NoConstraints
-		path = cgroupNoConstraintsPath(s.state.CgroupPath)
-		s.Logger().WithField("path", path).Debug("Deleting no constraints cgroup")
+		return s.cgroupMgr.Destroy()
 	}
 
+	cgroupSubsystems = V1NoConstraints
+	path = cgroupNoConstraintsPath(s.state.CgroupPath)
+	s.Logger().WithField("path", path).Debug("Deleting no constraints cgroup")
+
 	sandboxCgroups, err := cgroupsLoadFunc(cgroupSubsystems, cgroups.StaticPath(path))
 	if err == cgroups.ErrCgroupDeleted {
 		// cgroup already deleted
@@ -2049,60 +2094,27 @@ func (s *Sandbox) setupSandboxCgroup() error {
 
 	s.Logger().WithField("hasCRIContainerType", s.config.HasCRIContainerType).Debug("Setting sandbox cgroup")
 
-	s.state.CgroupPath, err = validCgroupPath(spec.Linux.CgroupsPath, s.config.SystemdCgroup)
+	s.state.CgroupPath, err = vccgroups.ValidCgroupPath(spec.Linux.CgroupsPath, s.config.SystemdCgroup)
 	if err != nil {
 		return fmt.Errorf("Invalid cgroup path: %v", err)
 	}
 
-	// Don't modify original resources, create a copy
-	resources := *spec.Linux.Resources
-	sandboxSpec := specs.Spec{
-		Linux: &specs.Linux{
-			Resources: &resources,
-		},
-	}
-
-	// kata should rely on the cgroup created and configured by
-	// container engine *only* if actual container was
-	// marked *explicitly* as sandbox through annotations.
-	if s.config.HasCRIContainerType {
-		// Do not change current cgroup configuration.
-		// Create a spec without constraints
-		sandboxSpec.Linux.Resources = &specs.LinuxResources{}
-	}
-
-	sandboxSpec.Linux.CgroupsPath = s.state.CgroupPath
-
-	// Remove this to improve device resource management, but first we need to fix some issues:
-	// - hypervisors will need access to following host's devices:
-	//   * /dev/kvm
-	//   * /dev/vhost-net
-	// - If devicemapper is the storage driver, hypervisor will need access to devicemapper devices:
-	//   * The list of cgroup devices MUST BE updated when a new container is created in the POD
-	sandboxSpec.Linux.Resources.Devices = []specs.LinuxDeviceCgroup{}
-
-	cmgr, err := newCgroupManager(s.config.Cgroups, s.state.CgroupPaths, &sandboxSpec)
-	if err != nil {
-		return fmt.Errorf("Could not create a new cgroup manager: %v", err)
-	}
-
 	runtimePid := os.Getpid()
-
 	// Add the runtime to the Kata sandbox cgroup
-	if err = cmgr.Apply(runtimePid); err != nil {
+	if err = s.cgroupMgr.Add(runtimePid); err != nil {
 		return fmt.Errorf("Could not add runtime PID %d to sandbox cgroup:  %v", runtimePid, err)
 	}
 
 	// `Apply` updates manager's Cgroups and CgroupPaths,
 	// they both need to be saved since are used to create
 	// or restore a cgroup managers.
-	if s.config.Cgroups, err = cmgr.GetCgroups(); err != nil {
+	if s.config.Cgroups, err = s.cgroupMgr.GetCgroups(); err != nil {
 		return fmt.Errorf("Could not get cgroup configuration:  %v", err)
 	}
 
-	s.state.CgroupPaths = cmgr.GetPaths()
+	s.state.CgroupPaths = s.cgroupMgr.GetPaths()
 
-	if err = cmgr.Set(&configs.Config{Cgroups: s.config.Cgroups}); err != nil {
+	if err = s.cgroupMgr.Apply(); err != nil {
 		return fmt.Errorf("Could not constrain cgroup: %v", err)
 	}
 

virtcontainers/sandbox_test.go

@@ -1514,6 +1514,7 @@ func TestSandbox_SetupSandboxCgroup(t *testing.T) {
 		}
 
 		t.Run(tt.name, func(t *testing.T) {
+			tt.s.createCgroupManager()
 			if err := tt.s.setupSandboxCgroup(); (err != nil) != tt.wantErr {
 				t.Errorf("Sandbox.SetupSandboxCgroupOnly() error = %v, wantErr %v", err, tt.wantErr)
 			}