gh-119213: Be More Careful About _PyArg_Parser.kwtuple Across Interpr…

…eters (gh-119331) _PyArg_Parser holds static global data generated for modules by Argument Clinic. The _PyArg_Parser.kwtuple field is a tuple object, even though it's stored within a static global. In some cases the tuple is statically allocated and thus it's okay that it gets shared by multiple interpreters. However, in other cases the tuple is set lazily, allocated from the heap using the active interprepreter at the point the tuple is needed. This is a problem once that interpreter is destroyed since _PyArg_Parser.kwtuple becomes at dangling pointer, leading to crashes. It isn't a problem if the tuple is allocated under the main interpreter, since its lifetime is bound to the lifetime of the runtime. The solution here is to temporarily switch to the main interpreter. The alternative would be to always statically allocate the tuple. This change also fixes a bug where only the most recent parser was added to the global linked list. (cherry picked from commit 8186500) Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
python · ericsnowcurrently · May 22, 2024 · May 22, 2024 · May 22, 2024 · 05a00aa3dc26ec917ddd8839918ce3f0713a42a0
commit 05a00aa3dc26ec917ddd8839918ce3f0713a42a0
diff --git a/Include/internal/pycore_global_objects_fini_generated.h b/Include/internal/pycore_global_objects_fini_generated.h
diff --git a/Include/internal/pycore_global_strings.h b/Include/internal/pycore_global_strings.h
@@ -710,6 +710,7 @@ struct _Py_global_strings {
        STRUCT_FOR_ID(sort)
        STRUCT_FOR_ID(source)
        STRUCT_FOR_ID(source_traceback)
+        STRUCT_FOR_ID(spam)
        STRUCT_FOR_ID(src)
        STRUCT_FOR_ID(src_dir_fd)
        STRUCT_FOR_ID(stacklevel)

diff --git a/Include/internal/pycore_runtime_init_generated.h b/Include/internal/pycore_runtime_init_generated.h
diff --git a/Include/internal/pycore_unicodeobject_generated.h b/Include/internal/pycore_unicodeobject_generated.h
diff --git a/Lib/test/test_capi/test_getargs.py b/Lib/test/test_capi/test_getargs.py
@@ -4,11 +4,17 @@
 import sys
 from test import support
 from test.support import import_helper
+from test.support import script_helper
 from test.support import warnings_helper
 # Skip this test if the _testcapi module isn't available.
 _testcapi = import_helper.import_module('_testcapi')
 from _testcapi import getargs_keywords, getargs_keyword_only

+try:
+    import _testinternalcapi
+except ImportError:
+    _testinternalcapi = NULL
+
 # > How about the following counterproposal. This also changes some of
 # > the other format codes to be a little more regular.
 # >
@@ -1346,6 +1352,33 @@ def test_nested_tuple(self):
                    "argument 1 must be sequence of length 1, not 0"):
                parse(((),), {}, '(' + f + ')', ['a'])

+    @unittest.skipIf(_testinternalcapi is None, 'needs _testinternalcapi')
+    def test_gh_119213(self):
+        rc, out, err = script_helper.assert_python_ok("-c", """if True:
+            from test import support
+            script = '''if True:
+                import _testinternalcapi
+                _testinternalcapi.gh_119213_getargs(spam='eggs')
+                '''
+            config = dict(
+                allow_fork=False,
+                allow_exec=False,
+                allow_threads=True,
+                allow_daemon_threads=False,
+                use_main_obmalloc=False,
+                gil=2,
+                check_multi_interp_extensions=True,
+            )
+            rc = support.run_in_subinterp_with_config(script, **config)
+            assert rc == 0
+
+            # The crash is different if the interpreter was not destroyed first.
+            #interpid = _testinternalcapi.create_interpreter()
+            #rc = _testinternalcapi.exec_interpreter(interpid, script)
+            #assert rc == 0
+            """)
+        self.assertEqual(rc, 0)
+

 if __name__ == "__main__":
    unittest.main()
diff --git a/Misc/NEWS.d/next/Core and Builtins/2024-05-21-11-27-14.gh-issue-119213.nxjxrt.rst b/Misc/NEWS.d/next/Core and Builtins/2024-05-21-11-27-14.gh-issue-119213.nxjxrt.rst
@@ -0,0 +1,3 @@
+Non-builtin modules built with argument clinic were crashing if used in a
+subinterpreter before the main interpreter.  The objects that were causing
+the problem by leaking between interpreters carelessly have been fixed.
diff --git a/Modules/_testinternalcapi.c b/Modules/_testinternalcapi.c
@@ -2006,6 +2006,25 @@ has_inline_values(PyObject *self, PyObject *obj)
    Py_RETURN_FALSE;
 }

+
+/*[clinic input]
+gh_119213_getargs
+
+    spam: object = None
+
+Test _PyArg_Parser.kwtuple
+[clinic start generated code]*/
+
+static PyObject *
+gh_119213_getargs_impl(PyObject *module, PyObject *spam)
+/*[clinic end generated code: output=d8d9c95d5b446802 input=65ef47511da80fc2]*/
+{
+    // It must never have been called in the main interprer
+    assert(!_Py_IsMainInterpreter(PyInterpreterState_Get()));
+    return Py_NewRef(spam);
+}
+
+
 static PyMethodDef module_functions[] = {
    {"get_configs", get_configs, METH_NOARGS},
    {"get_recursion_depth", get_recursion_depth, METH_NOARGS},
@@ -2096,6 +2115,7 @@ static PyMethodDef module_functions[] = {
 #ifdef _Py_TIER2
    {"uop_symbols_test", _Py_uop_symbols_test, METH_NOARGS},
 #endif
+    GH_119213_GETARGS_METHODDEF
    {NULL, NULL} /* sentinel */
 };


diff --git a/Modules/clinic/_testinternalcapi.c.h b/Modules/clinic/_testinternalcapi.c.h
diff --git a/Python/getargs.c b/Python/getargs.c
@@ -7,6 +7,7 @@
 #include "pycore_dict.h"          // _PyDict_HasOnlyStringKeys()
 #include "pycore_modsupport.h"    // export _PyArg_NoKeywords()
 #include "pycore_pylifecycle.h"   // _PyArg_Fini
+#include "pycore_pystate.h"       // _Py_IsMainInterpreter()
 #include "pycore_tuple.h"         // _PyTuple_ITEMS()
 #include "pycore_pyerrors.h"      // _Py_CalculateSuggestions()

@@ -1947,7 +1948,23 @@ _parser_init(void *arg)
    int owned;
    PyObject *kwtuple = parser->kwtuple;
    if (kwtuple == NULL) {
+        /* We may temporarily switch to the main interpreter to avoid
+         * creating a tuple that could outlive its owning interpreter. */
+        PyThreadState *save_tstate = NULL;
+        PyThreadState *temp_tstate = NULL;
+        if (!_Py_IsMainInterpreter(PyInterpreterState_Get())) {
+            temp_tstate = PyThreadState_New(_PyInterpreterState_Main());
+            if (temp_tstate == NULL) {
+                return -1;
+            }
+            save_tstate = PyThreadState_Swap(temp_tstate);
+        }
        kwtuple = new_kwtuple(keywords, len, pos);
+        if (temp_tstate != NULL) {
+            PyThreadState_Clear(temp_tstate);
+            (void)PyThreadState_Swap(save_tstate);
+            PyThreadState_Delete(temp_tstate);
+        }
        if (kwtuple == NULL) {
            return -1;
        }
@@ -1969,8 +1986,8 @@ _parser_init(void *arg)
    parser->next = _Py_atomic_load_ptr(&_PyRuntime.getargs.static_parsers);
    do {
        // compare-exchange updates parser->next on failure
-    } while (_Py_atomic_compare_exchange_ptr(&_PyRuntime.getargs.static_parsers,
-                                             &parser->next, parser));
+    } while (!_Py_atomic_compare_exchange_ptr(&_PyRuntime.getargs.static_parsers,
+                                              &parser->next, parser));
    return 0;
 }


@@ -51,6 +51,8 @@ def declare_parser(
            #endif
        """
    else:
+        # XXX Why do we not statically allocate the tuple
+        # for non-builtin modules?
        declarations = """
            #if defined(Py_BUILD_CORE) && !defined(Py_BUILD_CORE_MODULE)