gh-112529: Implement GC for free-threaded builds (#114262)

* gh-112529: Implement GC for free-threaded builds

This implements a mark and sweep GC for the free-threaded builds of
CPython. The implementation relies on mimalloc to find GC tracked
objects (i.e., "containers").

Include/internal/pycore_freelist.h

@@ -20,6 +20,7 @@ extern "C" {
 #  define PyFloat_MAXFREELIST 100
 #  define PyContext_MAXFREELIST 255
 # define _PyAsyncGen_MAXFREELIST 80
+#  define _PyObjectStackChunk_MAXFREELIST 4
 #else
 #  define PyTuple_NFREELISTS 0
 #  define PyTuple_MAXFREELIST 0
@@ -27,6 +28,7 @@ extern "C" {
 #  define PyFloat_MAXFREELIST 0
 #  define PyContext_MAXFREELIST 0
 #  define _PyAsyncGen_MAXFREELIST 0
+#  define _PyObjectStackChunk_MAXFREELIST 0
 #endif
 
 struct _Py_list_state {
@@ -93,6 +95,13 @@ struct _Py_async_gen_state {
 #endif
 };
 
+struct _PyObjectStackChunk;
+
+struct _Py_object_stack_state {
+    struct _PyObjectStackChunk *free_list;
+    Py_ssize_t numfree;
+};
+
 typedef struct _Py_freelist_state {
     struct _Py_float_state float_state;
     struct _Py_tuple_state tuple_state;
@@ -100,6 +109,7 @@ typedef struct _Py_freelist_state {
     struct _Py_slice_state slice_state;
     struct _Py_context_state context_state;
     struct _Py_async_gen_state async_gen_state;
+    struct _Py_object_stack_state object_stack_state;
 } _PyFreeListState;
 
 #ifdef __cplusplus

Include/internal/pycore_gc.h

@@ -37,10 +37,22 @@ static inline PyObject* _Py_FROM_GC(PyGC_Head *gc) {
 }
 
 
+/* Bit flags for ob_gc_bits (in Py_GIL_DISABLED builds) */
+#ifdef Py_GIL_DISABLED
+#  define _PyGC_BITS_TRACKED        (1)
+#  define _PyGC_BITS_FINALIZED      (2)
+#  define _PyGC_BITS_UNREACHABLE    (4)
+#  define _PyGC_BITS_FROZEN         (8)
+#endif
+
 /* True if the object is currently tracked by the GC. */
 static inline int _PyObject_GC_IS_TRACKED(PyObject *op) {
+#ifdef Py_GIL_DISABLED
+    return (op->ob_gc_bits & _PyGC_BITS_TRACKED) != 0;
+#else
     PyGC_Head *gc = _Py_AS_GC(op);
     return (gc->_gc_next != 0);
+#endif
 }
 #define _PyObject_GC_IS_TRACKED(op) _PyObject_GC_IS_TRACKED(_Py_CAST(PyObject*, op))
 
@@ -107,24 +119,29 @@ static inline void _PyGCHead_SET_PREV(PyGC_Head *gc, PyGC_Head *prev) {
     gc->_gc_prev = ((gc->_gc_prev & ~_PyGC_PREV_MASK) | uprev);
 }
 
-static inline int _PyGCHead_FINALIZED(PyGC_Head *gc) {
-    return ((gc->_gc_prev & _PyGC_PREV_MASK_FINALIZED) != 0);
-}
-static inline void _PyGCHead_SET_FINALIZED(PyGC_Head *gc) {
-    gc->_gc_prev |= _PyGC_PREV_MASK_FINALIZED;
-}
-
 static inline int _PyGC_FINALIZED(PyObject *op) {
+#ifdef Py_GIL_DISABLED
+    return (op->ob_gc_bits & _PyGC_BITS_FINALIZED) != 0;
+#else
     PyGC_Head *gc = _Py_AS_GC(op);
-    return _PyGCHead_FINALIZED(gc);
+    return ((gc->_gc_prev & _PyGC_PREV_MASK_FINALIZED) != 0);
+#endif
 }
 static inline void _PyGC_SET_FINALIZED(PyObject *op) {
+#ifdef Py_GIL_DISABLED
+    op->ob_gc_bits |= _PyGC_BITS_FINALIZED;
+#else
     PyGC_Head *gc = _Py_AS_GC(op);
-    _PyGCHead_SET_FINALIZED(gc);
+    gc->_gc_prev |= _PyGC_PREV_MASK_FINALIZED;
+#endif
 }
 static inline void _PyGC_CLEAR_FINALIZED(PyObject *op) {
+#ifdef Py_GIL_DISABLED
+    op->ob_gc_bits &= ~_PyGC_BITS_FINALIZED;
+#else
     PyGC_Head *gc = _Py_AS_GC(op);
     gc->_gc_prev &= ~_PyGC_PREV_MASK_FINALIZED;
+#endif
 }
 
 

Include/internal/pycore_object.h

@@ -322,6 +322,9 @@ static inline void _PyObject_GC_TRACK(
                           "object is in generation which is garbage collected",
                           filename, lineno, __func__);
 
+#ifdef Py_GIL_DISABLED
+    op->ob_gc_bits |= _PyGC_BITS_TRACKED;
+#else
     PyInterpreterState *interp = _PyInterpreterState_GET();
     PyGC_Head *generation0 = interp->gc.generation0;
     PyGC_Head *last = (PyGC_Head*)(generation0->_gc_prev);
@@ -329,6 +332,7 @@ static inline void _PyObject_GC_TRACK(
     _PyGCHead_SET_PREV(gc, last);
     _PyGCHead_SET_NEXT(gc, generation0);
     generation0->_gc_prev = (uintptr_t)gc;
+#endif
 }
 
 /* Tell the GC to stop tracking this object.
@@ -352,6 +356,9 @@ static inline void _PyObject_GC_UNTRACK(
                           "object not tracked by the garbage collector",
                           filename, lineno, __func__);
 
+#ifdef Py_GIL_DISABLED
+    op->ob_gc_bits &= ~_PyGC_BITS_TRACKED;
+#else
     PyGC_Head *gc = _Py_AS_GC(op);
     PyGC_Head *prev = _PyGCHead_PREV(gc);
     PyGC_Head *next = _PyGCHead_NEXT(gc);
@@ -359,6 +366,7 @@ static inline void _PyObject_GC_UNTRACK(
     _PyGCHead_SET_PREV(next, prev);
     gc->_gc_next = 0;
     gc->_gc_prev &= _PyGC_PREV_MASK_FINALIZED;
+#endif
 }
 
 // Macros to accept any type for the parameter, and to automatically pass

Include/internal/pycore_object_stack.h

@@ -0,0 +1,84 @@
+#ifndef Py_INTERNAL_OBJECT_STACK_H
+#define Py_INTERNAL_OBJECT_STACK_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef Py_BUILD_CORE
+#  error "this header requires Py_BUILD_CORE define"
+#endif
+
+// _PyObjectStack is a stack of Python objects implemented as a linked list of
+// fixed size buffers.
+
+// Chosen so that _PyObjectStackChunk is a power-of-two size.
+#define _Py_OBJECT_STACK_CHUNK_SIZE 254
+
+typedef struct _PyObjectStackChunk {
+    struct _PyObjectStackChunk *prev;
+    Py_ssize_t n;
+    PyObject *objs[_Py_OBJECT_STACK_CHUNK_SIZE];
+} _PyObjectStackChunk;
+
+typedef struct _PyObjectStack {
+    _PyObjectStackChunk *head;
+} _PyObjectStack;
+
+
+extern _PyObjectStackChunk *
+_PyObjectStackChunk_New(void);
+
+extern void
+_PyObjectStackChunk_Free(_PyObjectStackChunk *);
+
+extern void
+_PyObjectStackChunk_ClearFreeList(_PyFreeListState *state, int is_finalization);
+
+// Push an item onto the stack. Return -1 on allocation failure, 0 on success.
+static inline int
+_PyObjectStack_Push(_PyObjectStack *stack, PyObject *obj)
+{
+    _PyObjectStackChunk *buf = stack->head;
+    if (buf == NULL || buf->n == _Py_OBJECT_STACK_CHUNK_SIZE) {
+        buf = _PyObjectStackChunk_New();
+        if (buf == NULL) {
+            return -1;
+        }
+        buf->prev = stack->head;
+        buf->n = 0;
+        stack->head = buf;
+    }
+
+    assert(buf->n >= 0 && buf->n < _Py_OBJECT_STACK_CHUNK_SIZE);
+    buf->objs[buf->n] = obj;
+    buf->n++;
+    return 0;
+}
+
+// Pop the top item from the stack.  Return NULL if the stack is empty.
+static inline PyObject *
+_PyObjectStack_Pop(_PyObjectStack *stack)
+{
+    _PyObjectStackChunk *buf = stack->head;
+    if (buf == NULL) {
+        return NULL;
+    }
+    assert(buf->n > 0 && buf->n <= _Py_OBJECT_STACK_CHUNK_SIZE);
+    buf->n--;
+    PyObject *obj = buf->objs[buf->n];
+    if (buf->n == 0) {
+        stack->head = buf->prev;
+        _PyObjectStackChunk_Free(buf);
+    }
+    return obj;
+}
+
+// Remove all items from the stack
+extern void
+_PyObjectStack_Clear(_PyObjectStack *stack);
+
+#ifdef __cplusplus
+}
+#endif
+#endif  // !Py_INTERNAL_OBJECT_STACK_H

Include/object.h

@@ -212,7 +212,9 @@ struct _object {
 struct _PyMutex { uint8_t v; };
 
 struct _object {
-    uintptr_t ob_tid;           // thread id (or zero)
+    // ob_tid stores the thread id (or zero). It is also used by the GC to
+    // store linked lists and the computed "gc_refs" refcount.
+    uintptr_t ob_tid;
     uint16_t _padding;
     struct _PyMutex ob_mutex;   // per-object lock
     uint8_t ob_gc_bits;         // gc-related state

Lib/gzip.py

@@ -349,7 +349,7 @@ class GzipFile(_compression.BaseStream):
 
     def close(self):
         fileobj = self.fileobj
-        if fileobj is None:
+        if fileobj is None or self._buffer.closed:
             return
         try:
             if self.mode == WRITE:

Lib/test/test_gc.py

@@ -1,7 +1,7 @@
 import unittest
 import unittest.mock
 from test.support import (verbose, refcount_test,
-                          cpython_only, requires_subprocess)
+                          cpython_only, requires_subprocess, Py_GIL_DISABLED)
 from test.support.import_helper import import_module
 from test.support.os_helper import temp_dir, TESTFN, unlink
 from test.support.script_helper import assert_python_ok, make_script
@@ -815,6 +815,15 @@ class GCTests(unittest.TestCase):
         self.assertEqual(gc.get_freeze_count(), 0)
 
     def test_get_objects(self):
+        gc.collect()
+        l = []
+        l.append(l)
+        self.assertTrue(
+                any(l is element for element in gc.get_objects())
+        )
+
+    @unittest.skipIf(Py_GIL_DISABLED, 'need generational GC')
+    def test_get_objects_generations(self):
         gc.collect()
         l = []
         l.append(l)
@@ -1225,7 +1234,7 @@ class GCCallbackTests(unittest.TestCase):
         p.stderr.close()
         # Verify that stderr has a useful error message:
         self.assertRegex(stderr,
-            br'gc\.c:[0-9]+: gc_decref: Assertion "gc_get_refs\(g\) > 0" failed.')
+            br'gc.*\.c:[0-9]+: .*: Assertion "gc_get_refs\(.+\) .*" failed.')
         self.assertRegex(stderr,
             br'refcount is too small')
         # "address : 0x7fb5062efc18"

Lib/test/test_io.py

@@ -3652,10 +3652,8 @@ class TextIOWrapperTest(unittest.TestCase):
             codecs.lookup('utf-8')
 
             class C:
-                def __init__(self):
-                    self.buf = io.BytesIO()
                 def __del__(self):
-                    io.TextIOWrapper(self.buf, **{kwargs})
+                    io.TextIOWrapper(io.BytesIO(), **{kwargs})
                     print("ok")
             c = C()
             """.format(iomod=iomod, kwargs=kwargs)

Makefile.pre.in

@@ -439,6 +439,7 @@ PYTHON_OBJS=	\
 		Python/modsupport.o \
 		Python/mysnprintf.o \
 		Python/mystrtoul.o \
+		Python/object_stack.o \
 		Python/optimizer.o \
 		Python/optimizer_analysis.o \
 		Python/parking_lot.o \
@@ -1832,6 +1833,7 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_frame.h \
 		$(srcdir)/Include/internal/pycore_freelist.h \
 		$(srcdir)/Include/internal/pycore_function.h \
+		$(srcdir)/Include/internal/pycore_gc.h \
 		$(srcdir)/Include/internal/pycore_genobject.h \
 		$(srcdir)/Include/internal/pycore_getopt.h \
 		$(srcdir)/Include/internal/pycore_gil.h \
@@ -1853,6 +1855,7 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_namespace.h \
 		$(srcdir)/Include/internal/pycore_object.h \
 		$(srcdir)/Include/internal/pycore_object_alloc.h \
+		$(srcdir)/Include/internal/pycore_object_stack.h \
 		$(srcdir)/Include/internal/pycore_object_state.h \
 		$(srcdir)/Include/internal/pycore_obmalloc.h \
 		$(srcdir)/Include/internal/pycore_obmalloc_init.h \

Misc/NEWS.d/next/Core and Builtins/2024-01-18-20-20-37.gh-issue-112529.oVNvDG.rst

@@ -0,0 +1,3 @@
+The free-threaded build now has its own thread-safe GC implementation that
+uses mimalloc to find GC tracked objects. It is non-generational, unlike the
+existing GC implementation.

PCbuild/_freeze_module.vcxproj

@@ -230,6 +230,7 @@
     <ClCompile Include="..\Python\modsupport.c" />
     <ClCompile Include="..\Python\mysnprintf.c" />
     <ClCompile Include="..\Python\mystrtoul.c" />
+    <ClCompile Include="..\Python\object_stack.c" />
     <ClCompile Include="..\Python\optimizer.c" />
     <ClCompile Include="..\Python\optimizer_analysis.c" />
     <ClCompile Include="..\Python\parking_lot.c" />

PCbuild/_freeze_module.vcxproj.filters

@@ -289,6 +289,9 @@
     <ClCompile Include="..\Objects\object.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\Python\object_stack.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
     <ClCompile Include="..\Objects\obmalloc.c">
       <Filter>Source Files</Filter>
     </ClCompile>

PCbuild/pythoncore.vcxproj

@@ -591,6 +591,7 @@
     <ClCompile Include="..\Python\modsupport.c" />
     <ClCompile Include="..\Python\mysnprintf.c" />
     <ClCompile Include="..\Python\mystrtoul.c" />
+    <ClCompile Include="..\Python\object_stack.c" />
     <ClCompile Include="..\Python\optimizer.c" />
     <ClCompile Include="..\Python\optimizer_analysis.c" />
     <ClCompile Include="..\Python\parking_lot.c" />

PCbuild/pythoncore.vcxproj.filters

@@ -1355,6 +1355,9 @@
     <ClCompile Include="..\Python\mystrtoul.c">
       <Filter>Python</Filter>
     </ClCompile>
+    <ClCompile Include="..\Python\object_stack.c">
+      <Filter>Python</Filter>
+    </ClCompile>
     <ClCompile Include="..\Python\optimizer.c">
       <Filter>Python</Filter>
     </ClCompile>

Python/gc.c

@@ -15,6 +15,8 @@
 #include "pycore_weakref.h"       // _PyWeakref_ClearRef()
 #include "pydtrace.h"
 
+#ifndef Py_GIL_DISABLED
+
 typedef struct _gc_runtime_state GCState;
 
 #ifdef Py_DEBUG
@@ -964,10 +966,10 @@ finalize_garbage(PyThreadState *tstate, PyGC_Head *collectable)
         PyGC_Head *gc = GC_NEXT(collectable);
         PyObject *op = FROM_GC(gc);
         gc_list_move(gc, &seen);
-        if (!_PyGCHead_FINALIZED(gc) &&
+        if (!_PyGC_FINALIZED(op) &&
             (finalize = Py_TYPE(op)->tp_finalize) != NULL)
         {
-            _PyGCHead_SET_FINALIZED(gc);
+            _PyGC_SET_FINALIZED(op);
             Py_INCREF(op);
             finalize(op);
             assert(!_PyErr_Occurred(tstate));
@@ -1942,3 +1944,5 @@ PyUnstable_GC_VisitObjects(gcvisitobjects_t callback, void *arg)
 done:
     gcstate->enabled = origenstate;
 }
+
+#endif  // Py_GIL_DISABLED

Python/object_stack.c

@@ -0,0 +1,87 @@
+// Stack of Python objects
+
+#include "Python.h"
+#include "pycore_freelist.h"
+#include "pycore_pystate.h"
+#include "pycore_object_stack.h"
+
+extern _PyObjectStackChunk *_PyObjectStackChunk_New(void);
+extern void _PyObjectStackChunk_Free(_PyObjectStackChunk *);
+
+static struct _Py_object_stack_state *
+get_state(void)
+{
+    _PyFreeListState *state = _PyFreeListState_GET();
+    return &state->object_stack_state;
+}
+
+_PyObjectStackChunk *
+_PyObjectStackChunk_New(void)
+{
+    _PyObjectStackChunk *buf;
+    struct _Py_object_stack_state *state = get_state();
+    if (state->numfree > 0) {
+        buf = state->free_list;
+        state->free_list = buf->prev;
+        state->numfree--;
+    }
+    else {
+        // NOTE: we use PyMem_RawMalloc() here because this is used by the GC
+        // during mimalloc heap traversal. In that context, it is not safe to
+        // allocate mimalloc memory, such as via PyMem_Malloc().
+        buf = PyMem_RawMalloc(sizeof(_PyObjectStackChunk));
+        if (buf == NULL) {
+            return NULL;
+        }
+    }
+    buf->prev = NULL;
+    buf->n = 0;
+    return buf;
+}
+
+void
+_PyObjectStackChunk_Free(_PyObjectStackChunk *buf)
+{
+    assert(buf->n == 0);
+    struct _Py_object_stack_state *state = get_state();
+    if (state->numfree >= 0 &&
+        state->numfree < _PyObjectStackChunk_MAXFREELIST)
+    {
+        buf->prev = state->free_list;
+        state->free_list = buf;
+        state->numfree++;
+    }
+    else {
+        PyMem_RawFree(buf);
+    }
+}
+
+void
+_PyObjectStack_Clear(_PyObjectStack *queue)
+{
+    while (queue->head != NULL) {
+        _PyObjectStackChunk *buf = queue->head;
+        buf->n = 0;
+        queue->head = buf->prev;
+        _PyObjectStackChunk_Free(buf);
+    }
+}
+
+void
+_PyObjectStackChunk_ClearFreeList(_PyFreeListState *free_lists, int is_finalization)
+{
+    if (!is_finalization) {
+        // Ignore requests to clear the free list during GC. We use object
+        // stacks during GC, so emptying the free-list is counterproductive.
+        return;
+    }
+
+    struct _Py_object_stack_state *state = &free_lists->object_stack_state;
+    while (state->numfree > 0) {
+        _PyObjectStackChunk *buf = state->free_list;
+        state->free_list = buf->prev;
+        state->numfree--;
+        PyMem_RawFree(buf);
+    }
+    state->numfree = -1;
+}

Python/pystate.c

@@ -10,6 +10,7 @@
 #include "pycore_frame.h"
 #include "pycore_initconfig.h"    // _PyStatus_OK()
 #include "pycore_object.h"        // _PyType_InitCache()
+#include "pycore_object_stack.h"  // _PyObjectStackChunk_ClearFreeList()
 #include "pycore_parking_lot.h"   // _PyParkingLot_AfterFork()
 #include "pycore_pyerrors.h"      // _PyErr_Clear()
 #include "pycore_pylifecycle.h"   // _PyAST_Fini()
@@ -1468,6 +1469,7 @@ _Py_ClearFreeLists(_PyFreeListState *state, int is_finalization)
     _PyList_ClearFreeList(state, is_finalization);
     _PyContext_ClearFreeList(state, is_finalization);
     _PyAsyncGen_ClearFreeLists(state, is_finalization);
+    _PyObjectStackChunk_ClearFreeList(state, is_finalization);
 }
 
 void
@@ -2055,7 +2057,6 @@ start_the_world(struct _stoptheworld_state *stw)
     HEAD_LOCK(runtime);
     stw->requested = 0;
     stw->world_stopped = 0;
-    stw->requester = NULL;
     // Switch threads back to the detached state.
     PyInterpreterState *i;
     PyThreadState *t;
@@ -2066,6 +2067,7 @@ start_the_world(struct _stoptheworld_state *stw)
             _PyParkingLot_UnparkAll(&t->state);
         }
     }
+    stw->requester = NULL;
     HEAD_UNLOCK(runtime);
     if (stw->is_global) {
         _PyRWMutex_Unlock(&runtime->stoptheworld_mutex);