diff --git a/Lib/test/test_subprocess.py b/Lib/test/test_subprocess.py index 806a1e3fa303eb..7e04934c317a91 100644 --- a/Lib/test/test_subprocess.py +++ b/Lib/test/test_subprocess.py @@ -22,6 +22,7 @@ import sysconfig import select import shutil +import socket import threading import gc import textwrap @@ -1044,19 +1045,49 @@ def test_communicate_timeout_large_input(self): # On Windows, stdin writing must also honor the timeout rather than # blocking indefinitely when the pipe buffer fills. - # Input larger than typical pipe buffer (4-64KB on Windows) - input_data = b"x" * (128 * 1024) + input_data = b"x" * (128 * 1024) # > typical pipe buffer + + # Cross-platform wake mechanism: the slow reader connects to a + # loopback TCP socket and blocks in select() on it (capped at 9s + # as a safety net we don't expect to hit). After phase 1 raises + # TimeoutExpired, the parent sends a byte to release the child so + # it drains stdin. A socket (rather than a raw pipe) is required + # because Windows select() only supports sockets, not arbitrary + # file descriptors. + server = socket.create_server(('127.0.0.1', 0), backlog=1) + server.settimeout(10) # bound the accept() if the child fails to start + port = server.getsockname()[1] + # The child sends one byte (low byte of its PID) first so the parent + # can detect the rare case of an unrelated process on the same host + # connecting to our ephemeral port before our child does. A single + # byte gives 1/256 collision odds, which is plenty for flake-prevention. + slow_reader = ( + "import os, socket, sys, select; " + f"s = socket.create_connection(('127.0.0.1', {port}), timeout=9); " + "s.sendall(bytes([os.getpid() & 0xff])); " + "select.select([s], [], [], 9); " + "sys.stdout.buffer.write(sys.stdin.buffer.read())" + ) p = subprocess.Popen( - [sys.executable, "-c", - "import sys, time; " - "time.sleep(30); " # Don't read stdin for a long time - "sys.stdout.buffer.write(sys.stdin.buffer.read())"], + [sys.executable, "-c", slow_reader], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE) + conn = None try: + conn, _ = server.accept() + server.close() + server = None + + conn.settimeout(5) + peer_byte = conn.recv(1) + conn.settimeout(None) + self.assertEqual(peer_byte, bytes([p.pid & 0xff]), + f"loopback handshake byte {peer_byte!r} != " + f"low byte of child PID {p.pid} ({p.pid & 0xff:#x})") + timeout = 0.2 start = time.monotonic() try: @@ -1065,7 +1096,7 @@ def test_communicate_timeout_large_input(self): elapsed = time.monotonic() - start self.fail( f"TimeoutExpired not raised. communicate() completed in " - f"{elapsed:.2f}s, but subprocess sleeps for 30s. " + f"{elapsed:.2f}s, but slow reader stalls for up to 9s. " "Stdin writing blocked without enforcing timeout.") except subprocess.TimeoutExpired: elapsed = time.monotonic() - start @@ -1073,11 +1104,16 @@ def test_communicate_timeout_large_input(self): # Timeout should occur close to the specified timeout value, # not after waiting for the subprocess to finish sleeping. # Allow generous margin for slow CI, but must be well under - # the subprocess sleep time. + # the slow-reader's stall cap. self.assertLess(elapsed, 5.0, f"TimeoutExpired raised after {elapsed:.2f}s; expected ~{timeout}s. " "Stdin writing blocked without checking timeout.") + # Release the slow reader so it stops blocking and drains stdin. + conn.sendall(b'go') + conn.close() + conn = None + # After timeout, continue communication. The remaining input # should be sent and we should receive all data back. stdout, stderr = p.communicate() @@ -1087,6 +1123,10 @@ def test_communicate_timeout_large_input(self): f"Expected {len(input_data)} bytes output but got {len(stdout)}") self.assertEqual(stdout, input_data) finally: + if conn is not None: + conn.close() + if server is not None: + server.close() p.kill() p.wait()