pythongh-123358: Use _PyStackRef in LOAD_DEREF (pythongh-130064)

Concurrent accesses from multiple threads to the same `cell` object did not
scale well in the free-threaded build. Use `_PyStackRef` and optimistically
avoid locking to improve scaling.

With the locks around cell reads gone, some of the free threading tests were
prone to starvation: the readers were able to run in a tight loop and the
writer threads weren't scheduled frequently enough to make timely progress.
Adjust the tests to avoid this.

Co-authored-by: Donghee Na <donghee.na@python.org>

Author: colesbury

Include/internal/pycore_cell.h

@@ -2,6 +2,8 @@
 #define Py_INTERNAL_CELL_H
 
 #include "pycore_critical_section.h"
+#include "pycore_object.h"
+#include "pycore_stackref.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -19,7 +21,7 @@ PyCell_SwapTakeRef(PyCellObject *cell, PyObject *value)
     PyObject *old_value;
     Py_BEGIN_CRITICAL_SECTION(cell);
     old_value = cell->ob_ref;
-    cell->ob_ref = value;
+    FT_ATOMIC_STORE_PTR_RELEASE(cell->ob_ref, value);
     Py_END_CRITICAL_SECTION();
     return old_value;
 }
@@ -37,11 +39,36 @@ PyCell_GetRef(PyCellObject *cell)
 {
     PyObject *res;
     Py_BEGIN_CRITICAL_SECTION(cell);
+#ifdef Py_GIL_DISABLED
+    res = _Py_XNewRefWithLock(cell->ob_ref);
+#else
     res = Py_XNewRef(cell->ob_ref);
+#endif
     Py_END_CRITICAL_SECTION();
     return res;
 }
 
+static inline _PyStackRef
+_PyCell_GetStackRef(PyCellObject *cell)
+{
+    PyObject *value;
+#ifdef Py_GIL_DISABLED
+    value = _Py_atomic_load_ptr(&cell->ob_ref);
+    if (value == NULL) {
+        return PyStackRef_NULL;
+    }
+    _PyStackRef ref;
+    if (_Py_TryIncrefCompareStackRef(&cell->ob_ref, value, &ref)) {
+        return ref;
+    }
+#endif
+    value = PyCell_GetRef(cell);
+    if (value == NULL) {
+        return PyStackRef_NULL;
+    }
+    return PyStackRef_FromPyObjectSteal(value);
+}
+
 #ifdef __cplusplus
 }
 #endif

Include/internal/pycore_opcode_metadata.h

@@ -74,6 +74,7 @@ def writer_func(name):
             last = -1
             while True:
                 if CUR == last:
+                    time.sleep(0.001)
                     continue
                 elif CUR == OBJECT_COUNT:
                     break

Include/internal/pycore_uop_metadata.h

@@ -27,13 +27,13 @@ def set_func_annotation(f, b):
 
 @unittest.skipUnless(Py_GIL_DISABLED, "Enable only in FT build")
 class TestFTFuncAnnotations(TestCase):
-    NUM_THREADS = 8
+    NUM_THREADS = 4
 
     def test_concurrent_read(self):
         def f(x: int) -> int:
             return x + 1
 
-        for _ in range(100):
+        for _ in range(10):
             with concurrent.futures.ThreadPoolExecutor(max_workers=self.NUM_THREADS) as executor:
                 b = Barrier(self.NUM_THREADS)
                 futures = {executor.submit(get_func_annotation, f, b): i for i in range(self.NUM_THREADS)}
@@ -54,7 +54,7 @@ def test_concurrent_write(self):
         def bar(x: int, y: float) -> float:
             return y ** x
 
-        for _ in range(100):
+        for _ in range(10):
             with concurrent.futures.ThreadPoolExecutor(max_workers=self.NUM_THREADS) as executor:
                 b = Barrier(self.NUM_THREADS)
                 futures = {executor.submit(set_func_annotation, bar, b): i for i in range(self.NUM_THREADS)}

Lib/test/test_free_threading/test_dict.py

@@ -35,24 +35,30 @@ def mutator_thread():
             pass
 
     def test_get_referrers(self):
+        NUM_GC = 2
+        NUM_MUTATORS = 4
+
+        b = threading.Barrier(NUM_GC + NUM_MUTATORS)
         event = threading.Event()
 
         obj = MyObj()
 
         def gc_thread():
+            b.wait()
             for i in range(100):
                 o = gc.get_referrers(obj)
             event.set()
 
         def mutator_thread():
+            b.wait()
             while not event.is_set():
                 d1 = { "key": obj }
                 d2 = { "key": obj }
                 d3 = { "key": obj }
                 d4 = { "key": obj }
 
-        gcs = [Thread(target=gc_thread) for _ in range(2)]
-        mutators = [Thread(target=mutator_thread) for _ in range(4)]
+        gcs = [Thread(target=gc_thread) for _ in range(NUM_GC)]
+        mutators = [Thread(target=mutator_thread) for _ in range(NUM_MUTATORS)]
         with threading_helper.start_threads(gcs + mutators):
             pass
 

Lib/test/test_free_threading/test_func_annotations.py

@@ -20,22 +20,25 @@ class TestList(TestCase):
     def test_racing_iter_append(self):
         l = []
 
-        def writer_func():
+        barrier = Barrier(NTHREAD + 1)
+        def writer_func(l):
+            barrier.wait()
             for i in range(OBJECT_COUNT):
                 l.append(C(i + OBJECT_COUNT))
 
-        def reader_func():
+        def reader_func(l):
+            barrier.wait()
             while True:
                 count = len(l)
                 for i, x in enumerate(l):
                     self.assertEqual(x.v, i + OBJECT_COUNT)
                 if count == OBJECT_COUNT:
                     break
 
-        writer = Thread(target=writer_func)
+        writer = Thread(target=writer_func, args=(l,))
         readers = []
         for x in range(NTHREAD):
-            reader = Thread(target=reader_func)
+            reader = Thread(target=reader_func, args=(l,))
             readers.append(reader)
             reader.start()
 
@@ -47,11 +50,14 @@ def reader_func():
     def test_racing_iter_extend(self):
         l = []
 
+        barrier = Barrier(NTHREAD + 1)
         def writer_func():
+            barrier.wait()
             for i in range(OBJECT_COUNT):
                 l.extend([C(i + OBJECT_COUNT)])
 
         def reader_func():
+            barrier.wait()
             while True:
                 count = len(l)
                 for i, x in enumerate(l):

Lib/test/test_free_threading/test_gc.py

@@ -8,7 +8,7 @@
 
 from sys import monitoring
 from test.support import threading_helper
-from threading import Thread, _PyRLock
+from threading import Thread, _PyRLock, Barrier
 from unittest import TestCase
 
 
@@ -194,7 +194,9 @@ def during_threads(self):
 
 @threading_helper.requires_working_threading()
 class MonitoringMisc(MonitoringTestMixin, TestCase):
-    def register_callback(self):
+    def register_callback(self, barrier):
+        barrier.wait()
+
         def callback(*args):
             pass
 
@@ -206,8 +208,9 @@ def callback(*args):
     def test_register_callback(self):
         self.refs = []
         threads = []
-        for i in range(50):
-            t = Thread(target=self.register_callback)
+        barrier = Barrier(5)
+        for i in range(5):
+            t = Thread(target=self.register_callback, args=(barrier,))
             t.start()
             threads.append(t)
 

Lib/test/test_free_threading/test_list.py

@@ -45,26 +45,20 @@ def test_attr_cache_consistency(self):
         class C:
             x = 0
 
-        DONE = False
         def writer_func():
-            for i in range(3000):
+            for _ in range(3000):
                 C.x
                 C.x
                 C.x += 1
-            nonlocal DONE
-            DONE = True
 
         def reader_func():
-            while True:
+            for _ in range(3000):
                 # We should always see a greater value read from the type than the
                 # dictionary
                 a = C.__dict__['x']
                 b = C.x
                 self.assertGreaterEqual(b, a)
 
-                if DONE:
-                    break
-
         self.run_one(writer_func, reader_func)
 
     def test_attr_cache_consistency_subclass(self):
@@ -74,26 +68,20 @@ class C:
         class D(C):
             pass
 
-        DONE = False
         def writer_func():
-            for i in range(3000):
+            for _ in range(3000):
                 D.x
                 D.x
                 C.x += 1
-            nonlocal DONE
-            DONE = True
 
         def reader_func():
-            while True:
+            for _ in range(3000):
                 # We should always see a greater value read from the type than the
                 # dictionary
                 a = C.__dict__['x']
                 b = D.x
                 self.assertGreaterEqual(b, a)
 
-                if DONE:
-                    break
-
         self.run_one(writer_func, reader_func)
 
     def test___class___modification(self):
@@ -140,10 +128,18 @@ class ClassB(Base):
 
 
     def run_one(self, writer_func, reader_func):
-        writer = Thread(target=writer_func)
+        barrier = threading.Barrier(NTHREADS)
+
+        def wrap_target(target):
+            def wrapper():
+                barrier.wait()
+                target()
+            return wrapper
+
+        writer = Thread(target=wrap_target(writer_func))
         readers = []
-        for x in range(30):
-            reader = Thread(target=reader_func)
+        for x in range(NTHREADS - 1):
+            reader = Thread(target=wrap_target(reader_func))
             readers.append(reader)
             reader.start()
 

Lib/test/test_free_threading/test_monitoring.py

@@ -1822,12 +1822,11 @@ dummy_func(
 
         inst(LOAD_DEREF, ( -- value)) {
             PyCellObject *cell = (PyCellObject *)PyStackRef_AsPyObjectBorrow(GETLOCAL(oparg));
-            PyObject *value_o = PyCell_GetRef(cell);
-            if (value_o == NULL) {
+            value = _PyCell_GetStackRef(cell);
+            if (PyStackRef_IsNull(value)) {
                 _PyEval_FormatExcUnbound(tstate, _PyFrame_GetCode(frame), oparg);
                 ERROR_IF(true, error);
             }
-            value = PyStackRef_FromPyObjectSteal(value_o);
         }
 
         inst(STORE_DEREF, (v --)) {