gh-117783: Immortalize objects that use deferred reference counting (#118112)

Deferred reference counting is not fully implemented yet. As a temporary
measure, we immortalize objects that would use deferred reference
counting to avoid multi-threaded scaling bottlenecks.

This is only performed in the free-threaded build once the first
non-main thread is started. Additionally, some tests, including refleak
tests, suppress this behavior.

Include/internal/pycore_gc.h

@@ -312,6 +312,18 @@ struct _gc_runtime_state {
        collections, and are awaiting to undergo a full collection for
        the first time. */
     Py_ssize_t long_lived_pending;
+
+    /* gh-117783: Deferred reference counting is not fully implemented yet, so
+       as a temporary measure we treat objects using deferred referenence
+       counting as immortal. */
+    struct {
+        /* Immortalize objects instead of marking them as using deferred
+           reference counting. */
+        int enabled;
+
+        /* Set enabled=1 when the first background thread is created. */
+        int enable_on_thread_created;
+    } immortalize;
 #endif
 };
 
@@ -343,6 +355,11 @@ extern void _PyGC_ClearAllFreeLists(PyInterpreterState *interp);
 extern void _Py_ScheduleGC(PyThreadState *tstate);
 extern void _Py_RunGC(PyThreadState *tstate);
 
+#ifdef Py_GIL_DISABLED
+// gh-117783: Immortalize objects that use deferred reference counting
+extern void _PyGC_ImmortalizeDeferredObjects(PyInterpreterState *interp);
+#endif
+
 #ifdef __cplusplus
 }
 #endif

Lib/concurrent/futures/process.py

@@ -296,8 +296,9 @@ class _ExecutorManagerThread(threading.Thread):
         # if there is no pending work item.
         def weakref_cb(_,
                        thread_wakeup=self.thread_wakeup,
-                       shutdown_lock=self.shutdown_lock):
-            mp.util.debug('Executor collected: triggering callback for'
+                       shutdown_lock=self.shutdown_lock,
+                       mp_util_debug=mp.util.debug):
+            mp_util_debug('Executor collected: triggering callback for'
                           ' QueueManager wakeup')
             with shutdown_lock:
                 thread_wakeup.wakeup()

Lib/test/libregrtest/main.py

@@ -7,7 +7,8 @@ import sysconfig
 import time
 import trace
 
-from test.support import os_helper, MS_WINDOWS, flush_std_streams
+from test.support import (os_helper, MS_WINDOWS, flush_std_streams,
+                          suppress_immortalization)
 
 from .cmdline import _parse_args, Namespace
 from .findtests import findtests, split_test_packages, list_cases
@@ -526,7 +527,10 @@ class Regrtest:
             if self.num_workers:
                 self._run_tests_mp(runtests, self.num_workers)
             else:
-                self.run_tests_sequentially(runtests)
+                # gh-117783: don't immortalize deferred objects when tracking
+                # refleaks. Only releveant for the free-threaded build.
+                with suppress_immortalization(runtests.hunt_refleak):
+                    self.run_tests_sequentially(runtests)
 
             coverage = self.results.get_coverage_results()
             self.display_result(runtests)

Lib/test/libregrtest/single.py

@@ -303,7 +303,10 @@ def run_single_test(test_name: TestName, runtests: RunTests) -> TestResult:
     result = TestResult(test_name)
     pgo = runtests.pgo
     try:
-        _runtest(result, runtests)
+        # gh-117783: don't immortalize deferred objects when tracking
+        # refleaks. Only releveant for the free-threaded build.
+        with support.suppress_immortalization(runtests.hunt_refleak):
+            _runtest(result, runtests)
     except:
         if not pgo:
             msg = traceback.format_exc()

Lib/test/support/__init__.py

@@ -516,6 +516,25 @@ def has_no_debug_ranges():
 def requires_debug_ranges(reason='requires co_positions / debug_ranges'):
     return unittest.skipIf(has_no_debug_ranges(), reason)
 
+@contextlib.contextmanager
+def suppress_immortalization(suppress=True):
+    """Suppress immortalization of deferred objects."""
+    try:
+        import _testinternalcapi
+    except ImportError:
+        yield
+        return
+
+    if not suppress:
+        yield
+        return
+
+    old_values = _testinternalcapi.set_immortalize_deferred(False)
+    try:
+        yield
+    finally:
+        _testinternalcapi.set_immortalize_deferred(*old_values)
+
 MS_WINDOWS = (sys.platform == 'win32')
 
 # Is not actually used in tests, but is kept for compatibility.

Lib/test/test_capi/test_watchers.py

@@ -1,7 +1,9 @@
 import unittest
 
 from contextlib import contextmanager, ExitStack
-from test.support import catch_unraisable_exception, import_helper, gc_collect
+from test.support import (
+    catch_unraisable_exception, import_helper,
+    gc_collect, suppress_immortalization)
 
 
 # Skip this test if the _testcapi module isn't available.
@@ -382,6 +384,7 @@ class TestCodeObjectWatchers(unittest.TestCase):
         self.assertEqual(
             exp_destroyed_1, _testcapi.get_code_watcher_num_destroyed_events(1))
 
+    @suppress_immortalization()
     def test_code_object_events_dispatched(self):
         # verify that all counts are zero before any watchers are registered
         self.assert_event_counts(0, 0, 0, 0)
@@ -428,6 +431,7 @@ class TestCodeObjectWatchers(unittest.TestCase):
                 self.assertIsNone(cm.unraisable.object)
                 self.assertEqual(str(cm.unraisable.exc_value), "boom!")
 
+    @suppress_immortalization()
     def test_dealloc_error(self):
         co = _testcapi.code_newempty("test_watchers", "dummy0", 0)
         with self.code_watcher(2):

Lib/test/test_code.py

@@ -141,7 +141,8 @@ except ImportError:
     ctypes = None
 from test.support import (cpython_only,
                           check_impl_detail, requires_debug_ranges,
-                          gc_collect, Py_GIL_DISABLED)
+                          gc_collect, Py_GIL_DISABLED,
+                          suppress_immortalization)
 from test.support.script_helper import assert_python_ok
 from test.support import threading_helper, import_helper
 from test.support.bytecode_helper import instructions_with_positions
@@ -577,6 +578,7 @@ class CodeConstsTest(unittest.TestCase):
 
 class CodeWeakRefTest(unittest.TestCase):
 
+    @suppress_immortalization()
     def test_basic(self):
         # Create a code object in a clean environment so that we know we have
         # the only reference to it left.
@@ -827,6 +829,7 @@ if check_impl_detail(cpython=True) and ctypes is not None:
             self.assertEqual(GetExtra(f.__code__, FREE_INDEX+100,
                               ctypes.c_voidp(100)), 0)
 
+        @suppress_immortalization()
         def test_free_called(self):
             # Verify that the provided free function gets invoked
             # when the code object is cleaned up.
@@ -854,6 +857,7 @@ if check_impl_detail(cpython=True) and ctypes is not None:
             del f
 
         @threading_helper.requires_working_threading()
+        @suppress_immortalization()
         def test_free_different_thread(self):
             # Freeing a code object on a different thread then
             # where the co_extra was set should be safe.

Lib/test/test_functools.py

@@ -1833,6 +1833,7 @@ class TestLRU:
             return 1
         self.assertEqual(f.cache_parameters(), {'maxsize': 1000, "typed": True})
 
+    @support.suppress_immortalization()
     def test_lru_cache_weakrefable(self):
         @self.module.lru_cache
         def test_function(x):

Lib/test/test_weakref.py

@@ -13,7 +13,7 @@ import random
 import textwrap
 
 from test import support
-from test.support import script_helper, ALWAYS_EQ
+from test.support import script_helper, ALWAYS_EQ, suppress_immortalization
 from test.support import gc_collect
 from test.support import import_helper
 from test.support import threading_helper
@@ -651,6 +651,7 @@ class ReferencesTestCase(TestBase):
         # deallocation of c2.
         del c2
 
+    @suppress_immortalization()
     def test_callback_in_cycle(self):
         import gc
 
@@ -743,6 +744,7 @@ class ReferencesTestCase(TestBase):
         del c1, c2, C, D
         gc.collect()
 
+    @suppress_immortalization()
     def test_callback_in_cycle_resurrection(self):
         import gc
 
@@ -878,6 +880,7 @@ class ReferencesTestCase(TestBase):
         # No exception should be raised here
         gc.collect()
 
+    @suppress_immortalization()
     def test_classes(self):
         # Check that classes are weakrefable.
         class A(object):

Modules/_testinternalcapi.c

@@ -1957,6 +1957,27 @@ get_py_thread_id(PyObject *self, PyObject *Py_UNUSED(ignored))
 }
 #endif
 
+static PyObject *
+set_immortalize_deferred(PyObject *self, PyObject *value)
+{
+#ifdef Py_GIL_DISABLED
+    PyInterpreterState *interp = PyInterpreterState_Get();
+    int old_enabled = interp->gc.immortalize.enabled;
+    int old_enabled_on_thread = interp->gc.immortalize.enable_on_thread_created;
+    int enabled_on_thread = 0;
+    if (!PyArg_ParseTuple(value, "i|i",
+                          &interp->gc.immortalize.enabled,
+                          &enabled_on_thread))
+    {
+        return NULL;
+    }
+    interp->gc.immortalize.enable_on_thread_created = enabled_on_thread;
+    return Py_BuildValue("ii", old_enabled, old_enabled_on_thread);
+#else
+    return Py_BuildValue("OO", Py_False, Py_False);
+#endif
+}
+
 static PyObject *
 has_inline_values(PyObject *self, PyObject *obj)
 {
@@ -2050,6 +2071,7 @@ static PyMethodDef module_functions[] = {
 #ifdef Py_GIL_DISABLED
     {"py_thread_id", get_py_thread_id, METH_NOARGS},
 #endif
+    {"set_immortalize_deferred", set_immortalize_deferred, METH_VARARGS},
     {"uop_symbols_test", _Py_uop_symbols_test, METH_NOARGS},
     {NULL, NULL} /* sentinel */
 };

Objects/object.c

@@ -2430,6 +2430,13 @@ _PyObject_SetDeferredRefcount(PyObject *op)
     assert(PyType_IS_GC(Py_TYPE(op)));
     assert(_Py_IsOwnedByCurrentThread(op));
     assert(op->ob_ref_shared == 0);
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    if (interp->gc.immortalize.enabled) {
+        // gh-117696: immortalize objects instead of using deferred reference
+        // counting for now.
+        _Py_SetImmortal(op);
+        return;
+    }
     op->ob_gc_bits |= _PyGC_BITS_DEFERRED;
     op->ob_ref_local += 1;
     op->ob_ref_shared = _Py_REF_QUEUED;

Python/gc_free_threading.c

@@ -704,6 +704,12 @@ _PyGC_Init(PyInterpreterState *interp)
 {
     GCState *gcstate = &interp->gc;
 
+    if (_Py_IsMainInterpreter(interp)) {
+        // gh-117783: immortalize objects that would use deferred refcounting
+        // once the first non-main thread is created.
+        gcstate->immortalize.enable_on_thread_created = 1;
+    }
+
     gcstate->garbage = PyList_New(0);
     if (gcstate->garbage == NULL) {
         return _PyStatus_NO_MEMORY();
@@ -1781,6 +1787,30 @@ custom_visitor_wrapper(const mi_heap_t *heap, const mi_heap_area_t *area,
     return true;
 }
 
+// gh-117783: Immortalize objects that use deferred reference counting to
+// temporarily work around scaling bottlenecks.
+static bool
+immortalize_visitor(const mi_heap_t *heap, const mi_heap_area_t *area,
+                    void *block, size_t block_size, void *args)
+{
+    PyObject *op = op_from_block(block, args, false);
+    if (op != NULL && _PyObject_HasDeferredRefcount(op)) {
+        _Py_SetImmortal(op);
+        op->ob_gc_bits &= ~_PyGC_BITS_DEFERRED;
+    }
+    return true;
+}
+
+void
+_PyGC_ImmortalizeDeferredObjects(PyInterpreterState *interp)
+{
+    struct visitor_args args;
+    _PyEval_StopTheWorld(interp);
+    gc_visit_heaps(interp, &immortalize_visitor, &args);
+    interp->gc.immortalize.enabled = 1;
+    _PyEval_StartTheWorld(interp);
+}
+
 void
 PyUnstable_GC_VisitObjects(gcvisitobjects_t callback, void *arg)
 {

Python/pystate.c

@@ -1568,6 +1568,17 @@ new_threadstate(PyInterpreterState *interp, int whence)
         // Must be called with lock unlocked to avoid re-entrancy deadlock.
         PyMem_RawFree(new_tstate);
     }
+    else {
+#ifdef Py_GIL_DISABLED
+        if (interp->gc.immortalize.enable_on_thread_created &&
+            !interp->gc.immortalize.enabled)
+        {
+            // Immortalize objects marked as using deferred reference counting
+            // the first time a non-main thread is created.
+            _PyGC_ImmortalizeDeferredObjects(interp);
+        }
+#endif
+    }
 
 #ifdef Py_GIL_DISABLED
     // Must be called with lock unlocked to avoid lock ordering deadlocks.