pids: fix reading pids when running in a cgroup v1 container

OCI runtimes might mount the container cgroup at the cgroup root
itself, breaking the correspondence between what is displayed in
/proc/$PID/cgroup and the path under the cgroup root.

When the cgroup path doesn't exist, check if the process is still
alive and attempt to read the pids from the root cgroup.

Signed-off-by: Giuseppe Scrivano <[email protected]>

internal/proc/pids.go

@@ -69,7 +69,8 @@ func GetPIDsFromCgroup(pid string) ([]string, error) {
 // cgroup.
 func getPIDsFromCgroupV1(pid string) ([]string, error) {
 	// First, find the corresponding path to the PID cgroup.
-	f, err := os.Open(fmt.Sprintf("/proc/%s/cgroup", pid))
+	pidPath := fmt.Sprintf("/proc/%s/cgroup", pid)
+	f, err := os.Open(pidPath)
 	if err != nil {
 		return nil, err
 	}
@@ -95,7 +96,18 @@ func getPIDsFromCgroupV1(pid string) ([]string, error) {
 	// Second, extract the PIDs inside the cgroup.
 	f, err = os.Open(cgroupPath)
 	if err != nil {
-		return nil, err
+		if os.IsNotExist(err) {
+			// OCI runtimes might mount the container cgroup at the root, breaking what it showed
+			// in /proc/$PID/cgroup and the path.
+			// Check if the PID still exists to make sure the process is still alive.
+			if _, errStat := os.Stat(pidPath); errStat == nil {
+				cgroupPath = filepath.Join(cgroups.CgroupRoot, "pids", "cgroup.procs")
+				f, err = os.Open(cgroupPath)
+			}
+		}
+		if err != nil {
+			return nil, err
+		}
 	}
 	defer f.Close()