bpo-38630: Fix subprocess.Popen.send_signal() race condition (GH-16984)

On Unix, subprocess.Popen.send_signal() now polls the process status.
Polling reduces the risk of sending a signal to the wrong process if
the process completed, the Popen.returncode attribute is still None,
and the pid has been reassigned (recycled) to a new different
process.
This commit is contained in:
Victor Stinner 2020-01-15 17:38:55 +01:00 committed by GitHub
parent ed154c387e
commit e85a305503
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 57 additions and 3 deletions

View File

@ -774,6 +774,8 @@ Instances of the :class:`Popen` class have the following methods:
Sends the signal *signal* to the child.
Do nothing if the process completed.
.. note::
On Windows, SIGTERM is an alias for :meth:`terminate`. CTRL_C_EVENT and

View File

@ -2061,9 +2061,31 @@ class Popen(object):
def send_signal(self, sig):
"""Send a signal to the process."""
# Skip signalling a process that we know has already died.
if self.returncode is None:
os.kill(self.pid, sig)
# bpo-38630: Polling reduces the risk of sending a signal to the
# wrong process if the process completed, the Popen.returncode
# attribute is still None, and the pid has been reassigned
# (recycled) to a new different process. This race condition can
# happens in two cases.
#
# Case 1. Thread A calls Popen.poll(), thread B calls
# Popen.send_signal(). In thread A, waitpid() succeed and returns
# the exit status. Thread B calls kill() because poll() in thread A
# did not set returncode yet. Calling poll() in thread B prevents
# the race condition thanks to Popen._waitpid_lock.
#
# Case 2. waitpid(pid, 0) has been called directly, without
# using Popen methods: returncode is still None is this case.
# Calling Popen.poll() will set returncode to a default value,
# since waitpid() fails with ProcessLookupError.
self.poll()
if self.returncode is not None:
# Skip signalling a process that we know has already died.
return
# The race condition can still happen if the race condition
# described above happens between the returncode test
# and the kill() call.
os.kill(self.pid, sig)
def terminate(self):
"""Terminate the process with SIGTERM

View File

@ -3120,6 +3120,31 @@ class POSIXProcessTestCase(BaseTestCase):
self.assertEqual(returncode, -3)
def test_send_signal_race(self):
# bpo-38630: send_signal() must poll the process exit status to reduce
# the risk of sending the signal to the wrong process.
proc = subprocess.Popen(ZERO_RETURN_CMD)
# wait until the process completes without using the Popen APIs.
pid, status = os.waitpid(proc.pid, 0)
self.assertEqual(pid, proc.pid)
self.assertTrue(os.WIFEXITED(status), status)
self.assertEqual(os.WEXITSTATUS(status), 0)
# returncode is still None but the process completed.
self.assertIsNone(proc.returncode)
with mock.patch("os.kill") as mock_kill:
proc.send_signal(signal.SIGTERM)
# send_signal() didn't call os.kill() since the process already
# completed.
mock_kill.assert_not_called()
# Don't check the returncode value: the test reads the exit status,
# so Popen failed to read it and uses a default returncode instead.
self.assertIsNotNone(proc.returncode)
@unittest.skipUnless(mswindows, "Windows specific tests")
class Win32ProcessTestCase(BaseTestCase):

View File

@ -0,0 +1,5 @@
On Unix, :meth:`subprocess.Popen.send_signal` now polls the process status.
Polling reduces the risk of sending a signal to the wrong process if the
process completed, the :attr:`subprocess.Popen.returncode` attribute is still
``None``, and the pid has been reassigned (recycled) to a new different
process.