Really fix with the zombie test (hopefully)

So, instead of any timeouts for synchronization, we move to actual
events. We still have one (huge) timeout for pathological cases (just to
cleanup), but we don't sleep at all in normal usage.

The patch moves the zombie/cleaned up child creation out from setUp(),
and to on-demand by the functions. We remove all signal handler usage,
instead we rely on the fact that when a process is cleaned up, its FDs
are closed, so the read end of a pipe will get an event.

Reviewed-by: imsnah

test/ganeti.utils_unittest.py

@@ -32,6 +32,7 @@ import signal
 import socket
 import shutil
 import re
+import select
 
 import ganeti
 import testutils
@@ -45,36 +46,29 @@ from ganeti.utils import IsProcessAlive, RunCmd, \
 from ganeti.errors import LockError, UnitParseError, GenericError, \
      ProgrammerError
 
-def _ChildHandler(signal, stack):
-  global _ChildFlag
-  _ChildFlag = True
-
 
 class TestIsProcessAlive(unittest.TestCase):
   """Testing case for IsProcessAlive"""
 
-  def setUp(self):
-    global _ChildFlag
-    # create a (most probably) non-existing process-id
-    self.pid_non_existing = os.fork()
-    if self.pid_non_existing == 0:
-      os._exit(0)
-    elif self.pid_non_existing > 0:
-      os.waitpid(self.pid_non_existing, 0)
-    else:
-      raise SystemError("can't fork")
-    _ChildFlag = False
-    # Use _ChildHandler for SIGCHLD
-    self.chldOrig = signal.signal(signal.SIGCHLD, _ChildHandler)
+  def _CreateZombie(self):
     # create a zombie
-    self.pid_zombie = os.fork()
-    if self.pid_zombie == 0:
+    r_fd, w_fd = os.pipe()
+    pid_zombie = os.fork()
+    if pid_zombie == 0:
+      # explicit close of read, write end will be closed only due to exit
+      os.close(r_fd)
       os._exit(0)
-    elif self.pid_zombie < 0:
+    elif pid_zombie < 0:
       raise SystemError("can't fork")
-
-  def tearDown(self):
-    signal.signal(signal.SIGCHLD, self.chldOrig)
+    # parent: we close our end of the w_fd, so reads will error as
+    # soon as the OS cleans the child's filedescriptors on its exit
+    os.close(w_fd)
+    # wait for 60 seconds at max for the exit (pathological case, just
+    # so that the test doesn't hang indefinitely)
+    r_set, w_set, e_set = select.select([r_fd], [], [], 60)
+    if not r_set and not e_set:
+      self.fail("Timeout exceeded in zombie creation")
+    return pid_zombie
 
   def testExists(self):
     mypid = os.getpid()
@@ -82,23 +76,20 @@ class TestIsProcessAlive(unittest.TestCase):
                  "can't find myself running")
 
   def testZombie(self):
-    global _ChildFlag
-    timeout = 10
-
-    while not _ChildFlag:
-      if timeout >= 0:
-        time.sleep(0.2)
-        timeout -= 0.2
-      else:
-        self.fail("timed out waiting for child's signal")
-        break # not executed...
-
-    self.assert_(not IsProcessAlive(self.pid_zombie),
-                 "zombie not detected as zombie")
+    pid_zombie = self._CreateZombie()
+    is_zombie = not IsProcessAlive(pid_zombie)
+    self.assert_(is_zombie, "zombie not detected as zombie")
+    os.waitpid(pid_zombie, os.WNOHANG)
 
   def testNotExisting(self):
-    self.assert_(not IsProcessAlive(self.pid_non_existing),
-                 "noexisting process detected")
+    pid_non_existing = os.fork()
+    if pid_non_existing == 0:
+      os._exit(0)
+    elif pid_non_existing < 0:
+      raise SystemError("can't fork")
+    os.waitpid(pid_non_existing, 0)
+    self.assert_(not IsProcessAlive(pid_non_existing),
+                 "nonexisting process detected")
 
 
 class TestPidFileFunctions(unittest.TestCase):