pythongh-100227: Make the Global PyModuleDef Cache Safe for Isolated …

…Interpreters (pythongh-103084)

Sharing mutable (or non-immortal) objects between interpreters is generally not safe.  We can work around that but not easily. 
 There are two restrictions that are critical for objects that break interpreter isolation.

The first is that the object's state be guarded by a global lock.  For now the GIL meets this requirement, but a granular global lock is needed once we have a per-interpreter GIL.

The second restriction is that the object (and, for a container, its items) be deallocated/resized only when the interpreter in which it was allocated is the current one.  This is because every interpreter has (or will have, see pythongh-101660) its own object allocator.  Deallocating an object with a different allocator can cause crashes.

The dict for the cache of module defs is completely internal, which simplifies what we have to do to meet those requirements.  To do so, we do the following:

* add a mechanism for re-using a temporary thread state tied to the main interpreter in an arbitrary thread
   * add _PyRuntime.imports.extensions.main_tstate` 
   * add _PyThreadState_InitDetached() and _PyThreadState_ClearDetached() (pystate.c)
   * add _PyThreadState_BindDetached() and _PyThreadState_UnbindDetached() (pystate.c)
* make sure the cache dict (_PyRuntime.imports.extensions.dict) and its items are all owned by the main interpreter)
* add a placeholder using for a granular global lock

Note that the cache is only used for legacy extension modules and not for multi-phase init modules.

python#100227

Include/internal/pycore_import.h

@@ -14,13 +14,19 @@ struct _import_runtime_state {
        which is just about every time an extension module is imported.
        See PyInterpreterState.modules_by_index for more info. */
     Py_ssize_t last_module_index;
-    /* A dict mapping (filename, name) to PyModuleDef for modules.
-       Only legacy (single-phase init) extension modules are added
-       and only if they support multiple initialization (m_size >- 0)
-       or are imported in the main interpreter.
-       This is initialized lazily in _PyImport_FixupExtensionObject().
-       Modules are added there and looked up in _imp.find_extension(). */
-    PyObject *extensions;
+    struct {
+        /* A thread state tied to the main interpreter,
+           used exclusively for when the extensions dict is access/modified
+           from an arbitrary thread. */
+        PyThreadState main_tstate;
+        /* A dict mapping (filename, name) to PyModuleDef for modules.
+           Only legacy (single-phase init) extension modules are added
+           and only if they support multiple initialization (m_size >- 0)
+           or are imported in the main interpreter.
+           This is initialized lazily in _PyImport_FixupExtensionObject().
+           Modules are added there and looked up in _imp.find_extension(). */
+        PyObject *dict;
+    } extensions;
     /* Package context -- the full module name for package imports */
     const char * pkgcontext;
 };

Include/internal/pycore_pystate.h

@@ -127,6 +127,11 @@ PyAPI_FUNC(void) _PyThreadState_Init(
     PyThreadState *tstate);
 PyAPI_FUNC(void) _PyThreadState_DeleteExcept(PyThreadState *tstate);
 
+extern void _PyThreadState_InitDetached(PyThreadState *, PyInterpreterState *);
+extern void _PyThreadState_ClearDetached(PyThreadState *);
+extern void _PyThreadState_BindDetached(PyThreadState *);
+extern void _PyThreadState_UnbindDetached(PyThreadState *);
+
 
 static inline void
 _PyThreadState_UpdateTracingState(PyThreadState *tstate)

Include/internal/pycore_runtime_init.h

@@ -41,6 +41,11 @@ extern PyTypeObject _PyExc_MemoryError;
            in accordance with the specification. */ \
         .autoTSSkey = Py_tss_NEEDS_INIT, \
         .parser = _parser_runtime_state_INIT, \
+        .imports = { \
+            .extensions = { \
+                .main_tstate = _PyThreadState_INIT, \
+            }, \
+        }, \
         .ceval = { \
             .perf = _PyEval_RUNTIME_PERF_INIT, \
         }, \

Python/import.c

@@ -862,6 +862,18 @@ Generally, when multiple interpreters are involved, some of the above
 gets even messier.
 */
 
+static inline void
+extensions_lock_acquire(void)
+{
+    // XXX For now the GIL is sufficient.
+}
+
+static inline void
+extensions_lock_release(void)
+{
+    // XXX For now the GIL is sufficient.
+}
+
 /* Magic for extension modules (built-in as well as dynamically
    loaded).  To prevent initializing an extension module more than
    once, we keep a static dictionary 'extensions' keyed by the tuple
@@ -878,71 +890,170 @@ gets even messier.
    dictionary, to avoid loading shared libraries twice.
 */
 
+static void
+_extensions_cache_init(void)
+{
+    /* The runtime (i.e. main interpreter) must be initializing,
+       so we don't need to worry about the lock. */
+    _PyThreadState_InitDetached(&EXTENSIONS.main_tstate,
+                                _PyInterpreterState_Main());
+}
+
 static PyModuleDef *
 _extensions_cache_get(PyObject *filename, PyObject *name)
 {
-    PyObject *extensions = EXTENSIONS;
-    if (extensions == NULL) {
-        return NULL;
-    }
+    PyModuleDef *def = NULL;
+    extensions_lock_acquire();
+
     PyObject *key = PyTuple_Pack(2, filename, name);
     if (key == NULL) {
-        return NULL;
+        goto finally;
+    }
+
+    PyObject *extensions = EXTENSIONS.dict;
+    if (extensions == NULL) {
+        goto finally;
     }
-    PyModuleDef *def = (PyModuleDef *)PyDict_GetItemWithError(extensions, key);
-    Py_DECREF(key);
+    def = (PyModuleDef *)PyDict_GetItemWithError(extensions, key);
+
+finally:
+    Py_XDECREF(key);
+    extensions_lock_release();
     return def;
 }
 
 static int
 _extensions_cache_set(PyObject *filename, PyObject *name, PyModuleDef *def)
 {
-    PyObject *extensions = EXTENSIONS;
+    int res = -1;
+    PyThreadState *oldts = NULL;
+    extensions_lock_acquire();
+
+    /* Swap to the main interpreter, if necessary.  This matters if
+       the dict hasn't been created yet or if the item isn't in the
+       dict yet.  In both cases we must ensure the relevant objects
+       are created using the main interpreter. */
+    PyThreadState *main_tstate = &EXTENSIONS.main_tstate;
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    if (!_Py_IsMainInterpreter(interp)) {
+        _PyThreadState_BindDetached(main_tstate);
+        oldts = _PyThreadState_Swap(interp->runtime, main_tstate);
+        assert(!_Py_IsMainInterpreter(oldts->interp));
+
+        /* Make sure the name and filename objects are owned
+           by the main interpreter. */
+        name = PyUnicode_InternFromString(PyUnicode_AsUTF8(name));
+        assert(name != NULL);
+        filename = PyUnicode_InternFromString(PyUnicode_AsUTF8(filename));
+        assert(filename != NULL);
+    }
+
+    PyObject *key = PyTuple_Pack(2, filename, name);
+    if (key == NULL) {
+        goto finally;
+    }
+
+    PyObject *extensions = EXTENSIONS.dict;
     if (extensions == NULL) {
         extensions = PyDict_New();
         if (extensions == NULL) {
-            return -1;
+            goto finally;
         }
-        EXTENSIONS = extensions;
+        EXTENSIONS.dict = extensions;
     }
-    PyObject *key = PyTuple_Pack(2, filename, name);
-    if (key == NULL) {
-        return -1;
+
+    PyModuleDef *actual = (PyModuleDef *)PyDict_GetItemWithError(extensions, key);
+    if (PyErr_Occurred()) {
+        goto finally;
+    }
+    else if (actual != NULL) {
+        /* We expect it to be static, so it must be the same pointer. */
+        assert(def == actual);
+        res = 0;
+        goto finally;
     }
-    int res = PyDict_SetItem(extensions, key, (PyObject *)def);
-    Py_DECREF(key);
+
+    /* This might trigger a resize, which is why we must switch
+       to the main interpreter. */
+    res = PyDict_SetItem(extensions, key, (PyObject *)def);
     if (res < 0) {
-        return -1;
+        res = -1;
+        goto finally;
     }
-    return 0;
+    res = 0;
+
+finally:
+    if (oldts != NULL) {
+        _PyThreadState_Swap(interp->runtime, oldts);
+        _PyThreadState_UnbindDetached(main_tstate);
+        Py_DECREF(name);
+        Py_DECREF(filename);
+    }
+    Py_XDECREF(key);
+    extensions_lock_release();
+    return res;
 }
 
 static int
 _extensions_cache_delete(PyObject *filename, PyObject *name)
 {
-    PyObject *extensions = EXTENSIONS;
-    if (extensions == NULL) {
-        return 0;
-    }
+    int res = -1;
+    PyThreadState *oldts = NULL;
+    extensions_lock_acquire();
+
     PyObject *key = PyTuple_Pack(2, filename, name);
     if (key == NULL) {
-        return -1;
+        goto finally;
+    }
+
+    PyObject *extensions = EXTENSIONS.dict;
+    if (extensions == NULL) {
+        res = 0;
+        goto finally;
     }
+
+    PyModuleDef *actual = (PyModuleDef *)PyDict_GetItemWithError(extensions, key);
+    if (PyErr_Occurred()) {
+        goto finally;
+    }
+    else if (actual == NULL) {
+        /* It was already removed or never added. */
+        res = 0;
+        goto finally;
+    }
+
+    /* Swap to the main interpreter, if necessary. */
+    PyThreadState *main_tstate = &EXTENSIONS.main_tstate;
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    if (!_Py_IsMainInterpreter(interp)) {
+        _PyThreadState_BindDetached(main_tstate);
+        oldts = _PyThreadState_Swap(interp->runtime, main_tstate);
+        assert(!_Py_IsMainInterpreter(oldts->interp));
+    }
+
     if (PyDict_DelItem(extensions, key) < 0) {
-        if (!PyErr_ExceptionMatches(PyExc_KeyError)) {
-            Py_DECREF(key);
-            return -1;
-        }
-        PyErr_Clear();
+        goto finally;
     }
-    Py_DECREF(key);
-    return 0;
+    res = 0;
+
+finally:
+    if (oldts != NULL) {
+        _PyThreadState_Swap(interp->runtime, oldts);
+        _PyThreadState_UnbindDetached(main_tstate);
+    }
+    Py_XDECREF(key);
+    extensions_lock_release();
+    return res;
 }
 
 static void
 _extensions_cache_clear_all(void)
 {
-    Py_CLEAR(EXTENSIONS);
+    /* The runtime (i.e. main interpreter) must be finalizing,
+       so we don't need to worry about the lock. */
+    // XXX assert(_Py_IsMainInterpreter(_PyInterpreterState_GET()));
+    Py_CLEAR(EXTENSIONS.dict);
+    _PyThreadState_ClearDetached(&EXTENSIONS.main_tstate);
 }
 
 
@@ -2941,6 +3052,10 @@ _PyImport_Fini2(void)
 PyStatus
 _PyImport_InitCore(PyThreadState *tstate, PyObject *sysmod, int importlib)
 {
+    if (_Py_IsMainInterpreter(tstate->interp)) {
+        _extensions_cache_init();
+    }
+
     // XXX Initialize here: interp->modules and interp->import_func.
     // XXX Initialize here: sys.modules and sys.meta_path.