Fix benchexec when running in empty cgroup with cgroups v2

We document that one can start BenchExec in an otherwise cgroup.
However, this works only for runexec so far,
because we check that the current cgroup has exactly one process.
But benchexec starts the tool-info module in a separate process
if container mode is used, and so this check would fail.

We need to relax this and also allow child processes of us
in the given cgroup, and then later on move these child processes
into a different cgroup like our own process
(the existing code does this already).

benchexec/cgroupsv2.py

@@ -103,11 +103,11 @@ def initialize():
 
         cgroup = CgroupsV2.from_system()
 
-        if cgroup.path and list(cgroup.get_all_tasks()) == [os.getpid()]:
-            # If we are the only process, somebody prepared a cgroup for us. Use it.
-            # We might be able to relax this check and for example allow child processes,
-            # but then we would also have to move them to another cgroup,
-            # which might not be a good idea.
+        if cgroup.path and all(
+            util.is_child_process_of_us(pid) for pid in cgroup.get_all_tasks()
+        ):
+            # If we (and our subprocesses) are the only processes,
+            # somebody prepared a cgroup for us. Use it.
             logging.debug("BenchExec was started in its own cgroup: %s", cgroup)
 
         elif _create_systemd_scope_for_us() or _try_fallback_cgroup():
@@ -133,6 +133,8 @@ def initialize():
             logging.debug("Cgroup found, but cannot create child cgroups: %s", e)
             return CgroupsV2({})
 
+        # TODO: This moves our subprocesses if they are in the same cgroup.
+        # We should probably make this consistent and always move them.
         for pid in cgroup.get_all_tasks():
             child_cgroup.add_task(pid)
         assert child_cgroup.has_tasks()

benchexec/util.py

@@ -819,3 +819,24 @@ def check_msr():
         if all(os.access(f"/dev/cpu/{cpu}/msr", os.W_OK) for cpu in cpu_dirs):
             res["write"] = True
     return res
+
+
+def is_child_process_of_us(pid: int) -> bool:
+    """
+    Return if the given PID is a (transitive) child process of the current process.
+    Also returns true if the given PID is ours.
+    """
+    if pid == os.getpid():
+        return True
+
+    ppid = None
+    with open(f"/proc/{pid}/status") as status_file:
+        for line in status_file:
+            if line.startswith("PPid:"):
+                ppid = int(line.split(":", maxsplit=1)[1].strip())
+                break
+
+    if ppid:
+        return is_child_process_of_us(ppid)
+    else:
+        return False