add new pkg/systemd to contain podman systemd cgroup code

In podman we have code to move a process into a new systemd cgroup. This
code lived in the podman utils package. Because the new rootlessnetns
must call into that move this code to c/common.

Instead of dumping this again into a "util" package create a systemd
package which should have a better name. Also move the cgroup code
directly into pkg/cgroup. I am sure we can do some cleanup there in a
followup to prevent duplication.

Signed-off-by: Paul Holzinger <pholzing@redhat.com>

common/libnetwork/internal/rootlessnetns/netns_linux.go

@@ -15,6 +15,7 @@ import (
 	"github.com/containers/common/libnetwork/slirp4netns"
 	"github.com/containers/common/pkg/config"
 	"github.com/containers/common/pkg/netns"
+	"github.com/containers/common/pkg/systemd"
 	"github.com/containers/storage/pkg/homedir"
 	"github.com/containers/storage/pkg/lockfile"
 	"github.com/hashicorp/go-multierror"
@@ -163,15 +164,14 @@ func (n *Netns) setupSlirp4netns(nsPath string) error {
 		return wrapError("write slirp4netns pid file", err)
 	}
 
-	// FIXME
-	// if utils.RunsOnSystemd() {
-	// 	// move to systemd scope to prevent systemd from killing it
-	// 	err = utils.MoveRootlessNetnsSlirpProcessToUserSlice(res.Pid)
-	// 	if err != nil {
-	// 		// only log this, it is not fatal but can lead to issues when running podman inside systemd units
-	// 		logrus.Errorf("failed to move the rootless netns slirp4netns process to the systemd user.slice: %v", err)
-	// 	}
-	// }
+	if systemd.RunsOnSystemd() {
+		// move to systemd scope to prevent systemd from killing it
+		err = systemd.MoveRootlessNetnsSlirpProcessToUserSlice(res.Pid)
+		if err != nil {
+			// only log this, it is not fatal but can lead to issues when running podman inside systemd units
+			logrus.Errorf("failed to move the rootless netns slirp4netns process to the systemd user.slice: %v", err)
+		}
+	}
 
 	// build a new resolv.conf file which uses the slirp4netns dns server address
 	resolveIP, err := slirp4netns.GetDNS(res.Subnet)

common/pkg/cgroups/testdata/cgroup.other

@@ -0,0 +1 @@
+0::/other

common/pkg/cgroups/testdata/cgroup.root

@@ -0,0 +1 @@
+0::/

common/pkg/cgroups/utils_linux.go

@@ -4,6 +4,7 @@
 package cgroups
 
 import (
+	"bufio"
 	"bytes"
 	"errors"
 	"fmt"
@@ -11,6 +12,7 @@ import (
 	"path"
 	"path/filepath"
 	"strings"
+	"sync"
 
 	"github.com/opencontainers/runc/libcontainer/cgroups"
 	"github.com/opencontainers/runc/libcontainer/configs"
@@ -143,3 +145,171 @@ func SetBlkioThrottle(res *configs.Resources, cgroupPath string) error {
 	}
 	return nil
 }
+
+// Code below was moved from podman/utils/utils_supported.go and should properly better
+// integrated here as some parts may be redundant.
+
+func getCgroupProcess(procFile string, allowRoot bool) (string, error) {
+	f, err := os.Open(procFile)
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	scanner := bufio.NewScanner(f)
+	cgroup := ""
+	for scanner.Scan() {
+		line := scanner.Text()
+		parts := strings.SplitN(line, ":", 3)
+		if len(parts) != 3 {
+			return "", fmt.Errorf("cannot parse cgroup line %q", line)
+		}
+		if strings.HasPrefix(line, "0::") {
+			cgroup = line[3:]
+			break
+		}
+		if len(parts[2]) > len(cgroup) {
+			cgroup = parts[2]
+		}
+	}
+	if len(cgroup) == 0 || (!allowRoot && cgroup == "/") {
+		return "", fmt.Errorf("could not find cgroup mount in %q", procFile)
+	}
+	return cgroup, nil
+}
+
+// GetOwnCgroup returns the cgroup for the current process.
+func GetOwnCgroup() (string, error) {
+	return getCgroupProcess("/proc/self/cgroup", true)
+}
+
+func GetOwnCgroupDisallowRoot() (string, error) {
+	return getCgroupProcess("/proc/self/cgroup", false)
+}
+
+// GetCgroupProcess returns the cgroup for the specified process process.
+func GetCgroupProcess(pid int) (string, error) {
+	return getCgroupProcess(fmt.Sprintf("/proc/%d/cgroup", pid), true)
+}
+
+// MoveUnderCgroupSubtree moves the PID under a cgroup subtree.
+func MoveUnderCgroupSubtree(subtree string) error {
+	return MoveUnderCgroup("", subtree, nil)
+}
+
+// MoveUnderCgroup moves a group of processes to a new cgroup.
+// If cgroup is the empty string, then the current calling process cgroup is used.
+// If processes is empty, then the processes from the current cgroup are moved.
+func MoveUnderCgroup(cgroup, subtree string, processes []uint32) error {
+	procFile := "/proc/self/cgroup"
+	f, err := os.Open(procFile)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	unifiedMode, err := IsCgroup2UnifiedMode()
+	if err != nil {
+		return err
+	}
+
+	scanner := bufio.NewScanner(f)
+	for scanner.Scan() {
+		line := scanner.Text()
+		parts := strings.SplitN(line, ":", 3)
+		if len(parts) != 3 {
+			return fmt.Errorf("cannot parse cgroup line %q", line)
+		}
+
+		// root cgroup, skip it
+		if parts[2] == "/" && !(unifiedMode && parts[1] == "") {
+			continue
+		}
+
+		cgroupRoot := "/sys/fs/cgroup"
+		// Special case the unified mount on hybrid cgroup and named hierarchies.
+		// This works on Fedora 31, but we should really parse the mounts to see
+		// where the cgroup hierarchy is mounted.
+		if parts[1] == "" && !unifiedMode {
+			// If it is not using unified mode, the cgroup v2 hierarchy is
+			// usually mounted under /sys/fs/cgroup/unified
+			cgroupRoot = filepath.Join(cgroupRoot, "unified")
+
+			// Ignore the unified mount if it doesn't exist
+			if _, err := os.Stat(cgroupRoot); err != nil && os.IsNotExist(err) {
+				continue
+			}
+		} else if parts[1] != "" {
+			// Assume the controller is mounted at /sys/fs/cgroup/$CONTROLLER.
+			controller := strings.TrimPrefix(parts[1], "name=")
+			cgroupRoot = filepath.Join(cgroupRoot, controller)
+		}
+
+		parentCgroup := cgroup
+		if parentCgroup == "" {
+			parentCgroup = parts[2]
+		}
+		newCgroup := filepath.Join(cgroupRoot, parentCgroup, subtree)
+		if err := os.MkdirAll(newCgroup, 0o755); err != nil && !os.IsExist(err) {
+			return err
+		}
+
+		f, err := os.OpenFile(filepath.Join(newCgroup, "cgroup.procs"), os.O_RDWR, 0o755)
+		if err != nil {
+			return err
+		}
+		defer f.Close()
+
+		if len(processes) > 0 {
+			for _, pid := range processes {
+				if _, err := f.WriteString(fmt.Sprintf("%d\n", pid)); err != nil {
+					logrus.Debugf("Cannot move process %d to cgroup %q: %v", pid, newCgroup, err)
+				}
+			}
+		} else {
+			processesData, err := os.ReadFile(filepath.Join(cgroupRoot, parts[2], "cgroup.procs"))
+			if err != nil {
+				return err
+			}
+			for _, pid := range bytes.Split(processesData, []byte("\n")) {
+				if len(pid) == 0 {
+					continue
+				}
+				if _, err := f.Write(pid); err != nil {
+					logrus.Debugf("Cannot move process %s to cgroup %q: %v", string(pid), newCgroup, err)
+				}
+			}
+		}
+	}
+	return nil
+}
+
+var (
+	maybeMoveToSubCgroupSync    sync.Once
+	maybeMoveToSubCgroupSyncErr error
+)
+
+// MaybeMoveToSubCgroup moves the current process in a sub cgroup when
+// it is running in the root cgroup on a system that uses cgroupv2.
+func MaybeMoveToSubCgroup() error {
+	maybeMoveToSubCgroupSync.Do(func() {
+		unifiedMode, err := IsCgroup2UnifiedMode()
+		if err != nil {
+			maybeMoveToSubCgroupSyncErr = err
+			return
+		}
+		if !unifiedMode {
+			maybeMoveToSubCgroupSyncErr = nil
+			return
+		}
+		cgroup, err := GetOwnCgroup()
+		if err != nil {
+			maybeMoveToSubCgroupSyncErr = err
+			return
+		}
+		if cgroup == "/" {
+			maybeMoveToSubCgroupSyncErr = MoveUnderCgroupSubtree("init")
+		}
+	})
+	return maybeMoveToSubCgroupSyncErr
+}

common/pkg/cgroups/utils_linux_test.go

@@ -0,0 +1,23 @@
+package cgroups
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestCgroupProcess(t *testing.T) {
+	val, err := getCgroupProcess("testdata/cgroup.root", true)
+	assert.Nil(t, err)
+	assert.Equal(t, "/", val)
+
+	_, err = getCgroupProcess("testdata/cgroup.root", false)
+	assert.NotNil(t, err)
+
+	val, err = getCgroupProcess("testdata/cgroup.other", true)
+	assert.Nil(t, err)
+	assert.Equal(t, "/other", val)
+
+	_, err = getCgroupProcess("testdata/cgroup.empty", true)
+	assert.NotNil(t, err)
+}

common/pkg/systemd/systemd.go

@@ -0,0 +1,151 @@
+package systemd
+
+import (
+	"context"
+	"crypto/rand"
+	"fmt"
+	"os"
+	"strconv"
+	"sync"
+
+	"github.com/containers/common/pkg/cgroups"
+	"github.com/containers/storage/pkg/unshare"
+	systemdDbus "github.com/coreos/go-systemd/v22/dbus"
+	"github.com/godbus/dbus/v5"
+	"github.com/sirupsen/logrus"
+)
+
+var (
+	runsOnSystemdOnce sync.Once
+	runsOnSystemd     bool
+)
+
+// RunsOnSystemd returns whether the system is using systemd
+func RunsOnSystemd() bool {
+	runsOnSystemdOnce.Do(func() {
+		// per sd_booted(3), check for this dir
+		fd, err := os.Stat("/run/systemd/system")
+		runsOnSystemd = err == nil && fd.IsDir()
+	})
+	return runsOnSystemd
+}
+
+func moveProcessPIDFileToScope(pidPath, slice, scope string) error {
+	data, err := os.ReadFile(pidPath)
+	if err != nil {
+		// do not raise an error if the file doesn't exist
+		if os.IsNotExist(err) {
+			return nil
+		}
+		return fmt.Errorf("cannot read pid file: %w", err)
+	}
+	pid, err := strconv.ParseUint(string(data), 10, 0)
+	if err != nil {
+		return fmt.Errorf("cannot parse pid file %s: %w", pidPath, err)
+	}
+
+	return moveProcessToScope(int(pid), slice, scope)
+}
+
+func moveProcessToScope(pid int, slice, scope string) error {
+	err := RunUnderSystemdScope(pid, slice, scope)
+	// If the PID is not valid anymore, do not return an error.
+	if dbusErr, ok := err.(dbus.Error); ok {
+		if dbusErr.Name == "org.freedesktop.DBus.Error.UnixProcessIdUnknown" {
+			return nil
+		}
+	}
+	return err
+}
+
+// MoveRootlessNetnsSlirpProcessToUserSlice moves the slirp4netns process for the rootless netns
+// into a different scope so that systemd does not kill it with a container.
+func MoveRootlessNetnsSlirpProcessToUserSlice(pid int) error {
+	randBytes := make([]byte, 4)
+	_, err := rand.Read(randBytes)
+	if err != nil {
+		return err
+	}
+	return moveProcessToScope(pid, "user.slice", fmt.Sprintf("rootless-netns-%x.scope", randBytes))
+}
+
+// MovePauseProcessToScope moves the pause process used for rootless mode to keep the namespaces alive to
+// a separate scope.
+func MovePauseProcessToScope(pausePidPath string) {
+	var err error
+
+	for i := 0; i < 10; i++ {
+		randBytes := make([]byte, 4)
+		_, err = rand.Read(randBytes)
+		if err != nil {
+			logrus.Errorf("failed to read random bytes: %v", err)
+			continue
+		}
+		err = moveProcessPIDFileToScope(pausePidPath, "user.slice", fmt.Sprintf("podman-pause-%x.scope", randBytes))
+		if err == nil {
+			return
+		}
+	}
+
+	if err != nil {
+		unified, err2 := cgroups.IsCgroup2UnifiedMode()
+		if err2 != nil {
+			logrus.Warnf("Failed to detect if running with cgroup unified: %v", err)
+		}
+		if RunsOnSystemd() && unified {
+			logrus.Warnf("Failed to add pause process to systemd sandbox cgroup: %v", err)
+		} else {
+			logrus.Debugf("Failed to add pause process to systemd sandbox cgroup: %v", err)
+		}
+	}
+}
+
+// RunUnderSystemdScope adds the specified pid to a systemd scope
+func RunUnderSystemdScope(pid int, slice string, unitName string) error {
+	var properties []systemdDbus.Property
+	var conn *systemdDbus.Conn
+	var err error
+
+	if unshare.GetRootlessUID() != 0 {
+		conn, err = cgroups.UserConnection(unshare.GetRootlessUID())
+		if err != nil {
+			return err
+		}
+	} else {
+		conn, err = systemdDbus.NewWithContext(context.Background())
+		if err != nil {
+			return err
+		}
+	}
+	defer conn.Close()
+	properties = append(properties, systemdDbus.PropSlice(slice))
+	properties = append(properties, newProp("PIDs", []uint32{uint32(pid)}))
+	properties = append(properties, newProp("Delegate", true))
+	properties = append(properties, newProp("DefaultDependencies", false))
+	ch := make(chan string)
+	_, err = conn.StartTransientUnitContext(context.Background(), unitName, "replace", properties, ch)
+	if err != nil {
+		// On errors check if the cgroup already exists, if it does move the process there
+		if props, err := conn.GetUnitTypePropertiesContext(context.Background(), unitName, "Scope"); err == nil {
+			if cgroup, ok := props["ControlGroup"].(string); ok && cgroup != "" {
+				if err := cgroups.MoveUnderCgroup(cgroup, "", []uint32{uint32(pid)}); err == nil {
+					return nil
+				}
+				// On errors return the original error message we got from StartTransientUnit.
+			}
+		}
+		return err
+	}
+
+	// Block until job is started
+	<-ch
+
+	return nil
+}
+
+func newProp(name string, units interface{}) systemdDbus.Property {
+	return systemdDbus.Property{
+		Name:  name,
+		Value: dbus.MakeVariant(units),
+	}
+}