cgroups: kill processes when deleting a cgroup

if the cgroup cleanup fails with EBUSY, attempt to kill the
processes.

Related to: https://github.com/containers/podman/issues/14057

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>

common/pkg/cgroups/cgroups_supported.go

@@ -9,6 +9,7 @@ import (
 	"io/ioutil"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"sync"
 	"syscall"
@@ -96,6 +97,22 @@ func UserOwnsCurrentSystemdCgroup() (bool, error) {
 // It differs from os.RemoveAll as it doesn't attempt to unlink files.
 // On cgroupfs we are allowed only to rmdir empty directories.
 func rmDirRecursively(path string) error {
+	killProcesses := func(signal syscall.Signal) {
+		if signal == unix.SIGKILL {
+			if err := ioutil.WriteFile(filepath.Join(path, "cgroup.kill"), []byte("1"), 0600); err == nil {
+				return
+			}
+		}
+		// kill all the processes that are still part of the cgroup
+		if procs, err := ioutil.ReadFile(filepath.Join(path, "cgroup.procs")); err == nil {
+			for _, pidS := range strings.Split(string(procs), "\n") {
+				if pid, err := strconv.Atoi(pidS); err == nil {
+					_ = unix.Kill(pid, signal)
+				}
+			}
+		}
+	}
+
 	if err := os.Remove(path); err == nil || os.IsNotExist(err) {
 		return nil
 	}
@@ -118,8 +135,16 @@ func rmDirRecursively(path string) error {
 			return nil
 		}
 		if errors.Is(err, unix.EBUSY) {
-			// attempt up to 5 seconds if the cgroup is busy
-			if attempts < 500 {
+			// send a SIGTERM after 3 second
+			if attempts == 300 {
+				killProcesses(unix.SIGTERM)
+			}
+			// send SIGKILL after 8 seconds
+			if attempts == 800 {
+				killProcesses(unix.SIGKILL)
+			}
+			// give up after 10 seconds
+			if attempts < 1000 {
 				time.Sleep(time.Millisecond * 10)
 				attempts++
 				continue