[3.13] gh-117378: Fix multiprocessing forkserver preload sys.path inheritance. (GH-126538) (GH-126632)

gh-117378: Fix multiprocessing forkserver preload sys.path inheritance. (GH-126538)

`sys.path` was not properly being sent from the parent process when launching
the multiprocessing forkserver process to preload imports.  This bug has been
there since the forkserver start method was introduced in Python 3.4.  It was
always _supposed_ to inherit `sys.path` the same way the spawn method does.

Observable behavior change: A `''` value in `sys.path` will now be replaced in
the forkserver's `sys.path` with an absolute pathname
`os.path.abspath(os.getcwd())` saved at the time that `multiprocessing` was
imported in the parent process as it already was when using the spawn start
method. **This will only be observable during forkserver preload imports**.

The code invoked before calling things in another process already correctly sets `sys.path`.
Which is likely why this went unnoticed for so long as a mere performance issue in
some configurations.

A workaround for the bug on impacted Pythons is to set PYTHONPATH in the
environment before multiprocessing's forkserver process was started. Not perfect
as that is then inherited by other children, etc, but likely good enough for many
people's purposes.

(cherry picked from commit 9d08423)

Co-authored-by: Gregory P. Smith <[email protected]>
Co-authored-by: Serhiy Storchaka <[email protected]>

Author: 3 people

Lib/multiprocessing/forkserver.py

@@ -168,6 +168,8 @@ def ensure_running(self):
 def main(listener_fd, alive_r, preload, main_path=None, sys_path=None):
     '''Run forkserver.'''
     if preload:
+        if sys_path is not None:
+            sys.path[:] = sys_path
         if '__main__' in preload and main_path is not None:
             process.current_process()._inheriting = True
             try:

Lib/test/_test_multiprocessing.py

@@ -12,6 +12,7 @@
 import sys
 import os
 import gc
+import importlib
 import errno
 import functools
 import signal
@@ -20,8 +21,10 @@
 import socket
 import random
 import logging
+import shutil
 import subprocess
 import struct
+import tempfile
 import operator
 import pickle
 import weakref
@@ -6343,6 +6346,81 @@ def test_atexit(self):
                 self.assertEqual(f.read(), 'deadbeef')
 
 
+class _TestSpawnedSysPath(BaseTestCase):
+    """Test that sys.path is setup in forkserver and spawn processes."""
+
+    ALLOWED_TYPES = ('processes',)
+
+    def setUp(self):
+        self._orig_sys_path = list(sys.path)
+        self._temp_dir = tempfile.mkdtemp(prefix="test_sys_path-")
+        self._mod_name = "unique_test_mod"
+        module_path = os.path.join(self._temp_dir, f"{self._mod_name}.py")
+        with open(module_path, "w", encoding="utf-8") as mod:
+            mod.write("# A simple test module\n")
+        sys.path[:] = [p for p in sys.path if p]  # remove any existing ""s
+        sys.path.insert(0, self._temp_dir)
+        sys.path.insert(0, "")  # Replaced with an abspath in child.
+        try:
+            self._ctx_forkserver = multiprocessing.get_context("forkserver")
+        except ValueError:
+            self._ctx_forkserver = None
+        self._ctx_spawn = multiprocessing.get_context("spawn")
+
+    def tearDown(self):
+        sys.path[:] = self._orig_sys_path
+        shutil.rmtree(self._temp_dir, ignore_errors=True)
+
+    @staticmethod
+    def enq_imported_module_names(queue):
+        queue.put(tuple(sys.modules))
+
+    def test_forkserver_preload_imports_sys_path(self):
+        ctx = self._ctx_forkserver
+        if not ctx:
+            self.skipTest("requires forkserver start method.")
+        self.assertNotIn(self._mod_name, sys.modules)
+        multiprocessing.forkserver._forkserver._stop()  # Must be fresh.
+        ctx.set_forkserver_preload(
+            ["test.test_multiprocessing_forkserver", self._mod_name])
+        q = ctx.Queue()
+        proc = ctx.Process(target=self.enq_imported_module_names, args=(q,))
+        proc.start()
+        proc.join()
+        child_imported_modules = q.get()
+        q.close()
+        self.assertIn(self._mod_name, child_imported_modules)
+
+    @staticmethod
+    def enq_sys_path_and_import(queue, mod_name):
+        queue.put(sys.path)
+        try:
+            importlib.import_module(mod_name)
+        except ImportError as exc:
+            queue.put(exc)
+        else:
+            queue.put(None)
+
+    def test_child_sys_path(self):
+        for ctx in (self._ctx_spawn, self._ctx_forkserver):
+            if not ctx:
+                continue
+            with self.subTest(f"{ctx.get_start_method()} start method"):
+                q = ctx.Queue()
+                proc = ctx.Process(target=self.enq_sys_path_and_import,
+                                   args=(q, self._mod_name))
+                proc.start()
+                proc.join()
+                child_sys_path = q.get()
+                import_error = q.get()
+                q.close()
+                self.assertNotIn("", child_sys_path)  # replaced by an abspath
+                self.assertIn(self._temp_dir, child_sys_path)  # our addition
+                # ignore the first element, it is the absolute "" replacement
+                self.assertEqual(child_sys_path[1:], sys.path[1:])
+                self.assertIsNone(import_error, msg=f"child could not import {self._mod_name}")
+
+
 class MiscTestCase(unittest.TestCase):
     def test__all__(self):
         # Just make sure names in not_exported are excluded

Misc/NEWS.d/next/Library/2024-11-07-01-40-11.gh-issue-117378.o9O5uM.rst

@@ -0,0 +1,17 @@
+Fixed the :mod:`multiprocessing` ``"forkserver"`` start method forkserver
+process to correctly inherit the parent's :data:`sys.path` during the importing
+of :func:`multiprocessing.set_forkserver_preload` modules in the same manner as
+:data:`sys.path` is configured in workers before executing work items.
+
+This bug caused some forkserver module preloading to silently fail to preload.
+This manifested as a performance degration in child processes when the
+``sys.path`` was required due to additional repeated work in every worker.
+
+It could also have a side effect of ``""`` remaining in :data:`sys.path` during
+forkserver preload imports instead of the absolute path from :func:`os.getcwd`
+at multiprocessing import time used in the worker ``sys.path``.
+
+Potentially leading to incorrect imports from the wrong location during
+preload.  We are unaware of that actually happening.  The issue was discovered
+by someone observing unexpected preload performance gains.
+