Fix concurrency consistency for internals_pp_manager under multiple-interpreters

include/pybind11/detail/common.h

@@ -441,12 +441,11 @@ Note that this is run once for each (sub-)interpreter the module is imported int
 possibly concurrently.  The PyModuleDef is allowed to be static, but the PyObject* resulting from
 PyModuleDef_Init should be treated like any other PyObject (so not shared across interpreters).
  */
-#define PYBIND11_MODULE_PYINIT(name, pre_init, ...)                                               \
+#define PYBIND11_MODULE_PYINIT(name, ...)                                                         \
     static int PYBIND11_CONCAT(pybind11_exec_, name)(PyObject *);                                 \
     PYBIND11_PLUGIN_IMPL(name) {                                                                  \
         PYBIND11_CHECK_PYTHON_VERSION                                                             \
-        pre_init;                                                                                 \
-        PYBIND11_ENSURE_INTERNALS_READY                                                           \
+        pybind11::detail::ensure_internals();                                                     \
         static ::pybind11::detail::slots_array mod_def_slots = ::pybind11::detail::init_slots(    \
             &PYBIND11_CONCAT(pybind11_exec_, name), ##__VA_ARGS__);                               \
         static PyModuleDef def{/* m_base */ PyModuleDef_HEAD_INIT,                                \
@@ -465,6 +464,7 @@ PyModuleDef_Init should be treated like any other PyObject (so not shared across
     static void PYBIND11_CONCAT(pybind11_init_, name)(::pybind11::module_ &);                     \
     int PYBIND11_CONCAT(pybind11_exec_, name)(PyObject * pm) {                                    \
         try {                                                                                     \
+            pybind11::detail::ensure_internals();                                                 \
             auto m = pybind11::reinterpret_borrow<::pybind11::module_>(pm);                       \
             if (!pybind11::detail::get_cached_module(m.attr("__spec__").attr("name"))) {          \
                 PYBIND11_CONCAT(pybind11_init_, name)(m);                                         \
@@ -518,8 +518,7 @@ PyModuleDef_Init should be treated like any other PyObject (so not shared across
 
 \endrst */
 #define PYBIND11_MODULE(name, variable, ...)                                                      \
-    PYBIND11_MODULE_PYINIT(                                                                       \
-        name, (pybind11::detail::get_num_interpreters_seen() += 1), ##__VA_ARGS__)                \
+    PYBIND11_MODULE_PYINIT(name, ##__VA_ARGS__)                                                   \
     PYBIND11_MODULE_EXEC(name, variable)
 
 // pop gnu-zero-variadic-macro-arguments

include/pybind11/detail/internals.h

@@ -418,11 +418,12 @@ inline PyThreadState *get_thread_state_unchecked() {
 #endif
 }
 
-/// We use this counter to figure out if there are or have been multiple subinterpreters active at
-/// any point. This must never decrease while any interpreter may be running in any thread!
-inline std::atomic<int64_t> &get_num_interpreters_seen() {
-    static std::atomic<int64_t> counter(0);
-    return counter;
+/// We use this to figure out if there are or have been multiple subinterpreters active at any
+/// point. This must never go from true to false while any interpreter may be running in any
+/// thread!
+inline std::atomic<bool> &get_multiple_interpreters_seen() {
+    static std::atomic<bool> multi(false);
+    return multi;
 }
 
 template <class T,
@@ -649,7 +650,7 @@ class internals_pp_manager {
     /// acquire the GIL. Will never return nullptr.
     std::unique_ptr<InternalsType> *get_pp() {
 #ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
-        if (get_num_interpreters_seen() > 1) {
+        if (get_multiple_interpreters_seen()) {
             // Whenever the interpreter changes on the current thread we need to invalidate the
             // internals_pp so that it can be pulled from the interpreter's state dict.  That is
             // slow, so we use the current PyThreadState to check if it is necessary.
@@ -675,7 +676,7 @@ class internals_pp_manager {
     /// Drop all the references we're currently holding.
     void unref() {
 #ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
-        if (get_num_interpreters_seen() > 1) {
+        if (get_multiple_interpreters_seen()) {
             last_istate_tls() = nullptr;
             internals_p_tls() = nullptr;
             return;
@@ -686,7 +687,7 @@ class internals_pp_manager {
 
     void destroy() {
 #ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
-        if (get_num_interpreters_seen() > 1) {
+        if (get_multiple_interpreters_seen()) {
             auto *tstate = get_thread_state_unchecked();
             // this could be called without an active interpreter, just use what was cached
             if (!tstate || tstate->interp == last_istate_tls()) {
@@ -791,6 +792,16 @@ PYBIND11_NOINLINE internals &get_internals() {
     return *internals_ptr;
 }
 
+inline void ensure_internals() {
+    pybind11::detail::get_internals_pp_manager().unref();
+#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
+    if (PyInterpreterState_Get() != PyInterpreterState_Main()) {
+        get_multiple_interpreters_seen() = true;
+    }
+#endif
+    pybind11::detail::get_internals();
+}
+
 inline internals_pp_manager<local_internals> &get_local_internals_pp_manager() {
     // Use the address of this static itself as part of the key, so that the value is uniquely tied
     // to where the module is loaded in memory

include/pybind11/embed.h

@@ -58,7 +58,7 @@
 PYBIND11_WARNING_PUSH
 PYBIND11_WARNING_DISABLE_CLANG("-Wgnu-zero-variadic-macro-arguments")
 #define PYBIND11_EMBEDDED_MODULE(name, variable, ...)                                             \
-    PYBIND11_MODULE_PYINIT(name, {}, ##__VA_ARGS__)                                               \
+    PYBIND11_MODULE_PYINIT(name, ##__VA_ARGS__)                                                   \
     ::pybind11::detail::embedded_module PYBIND11_CONCAT(pybind11_module_, name)(                  \
         PYBIND11_TOSTRING(name), PYBIND11_CONCAT(PyInit_, name));                                 \
     PYBIND11_MODULE_EXEC(name, variable)
@@ -202,7 +202,7 @@ inline void initialize_interpreter(bool init_signal_handlers = true,
 #endif
 
     // There is exactly one interpreter alive currently.
-    detail::get_num_interpreters_seen() = 1;
+    detail::get_multiple_interpreters_seen() = false;
 }
 
 /** \rst
@@ -242,12 +242,12 @@ inline void initialize_interpreter(bool init_signal_handlers = true,
  \endrst */
 inline void finalize_interpreter() {
     // get rid of any thread-local interpreter cache that currently exists
-    if (detail::get_num_interpreters_seen() > 1) {
+    if (detail::get_multiple_interpreters_seen()) {
         detail::get_internals_pp_manager().unref();
         detail::get_local_internals_pp_manager().unref();
 
-        // We know there can be no other interpreter alive now, so we can lower the count
-        detail::get_num_interpreters_seen() = 1;
+        // We know there can be no other interpreter alive now
+        detail::get_multiple_interpreters_seen() = false;
     }
 
     // Re-fetch the internals pointer-to-pointer (but not the internals itself, which might not
@@ -265,8 +265,8 @@ inline void finalize_interpreter() {
     // avoid undefined behaviors when initializing another interpreter
     detail::get_local_internals_pp_manager().destroy();
 
-    // We know there is no interpreter alive now, so we can reset the count
-    detail::get_num_interpreters_seen() = 0;
+    // We know there is no interpreter alive now, so we can reset the multi-flag
+    detail::get_multiple_interpreters_seen() = false;
 }
 
 /** \rst

include/pybind11/gil_safe_call_once.h

@@ -232,7 +232,7 @@ class gil_safe_call_once_and_store {
     // Indicator of fast path for single-interpreter case.
     bool is_last_storage_valid() const {
         return is_initialized_by_at_least_one_interpreter_
-               && detail::get_num_interpreters_seen() == 1;
+               && !detail::get_multiple_interpreters_seen();
     }
 
     // Get the unique key for this storage instance in the interpreter's state dict.

include/pybind11/subinterpreter.h

@@ -109,7 +109,7 @@ class subinterpreter {
 
             // upon success, the new interpreter is activated in this thread
             result.istate_ = result.creation_tstate_->interp;
-            detail::get_num_interpreters_seen() += 1; // there are now many interpreters
+            detail::get_multiple_interpreters_seen() = true; // there are now many interpreters
             detail::get_internals(); // initialize internals.tstate, amongst other things...
 
             // In 3.13+ this state should be deleted right away, and the memory will be reused for

tests/mod_per_interpreter_gil_with_singleton.cpp

@@ -9,6 +9,7 @@ namespace py = pybind11;
 #    include <pybind11/native_enum.h>
 #endif
 
+namespace {
 // A singleton class that holds references to certain Python objects
 // This singleton is per-interpreter using gil_safe_call_once_and_store
 class MySingleton {
@@ -95,6 +96,7 @@ enum class MyEnum : int {
     TWO = 2,
     THREE = 3,
 };
+} // namespace
 
 PYBIND11_MODULE(mod_per_interpreter_gil_with_singleton,
                 m,

tests/test_multiple_interpreters.py

@@ -222,15 +222,27 @@
 
         objects = m.get_objects_in_singleton()
         expected = [
-            type(None),
-            tuple,
-            list,
-            dict,
-            collections.OrderedDict,
-            collections.defaultdict,
-            collections.deque,
+            type(None),               # static type: shared between interpreters
+            tuple,                    # static type: shared between interpreters
+            list,                     # static type: shared between interpreters
+            dict,                     # static type: shared between interpreters
+            collections.OrderedDict,  # static type: shared between interpreters
+            collections.defaultdict,  # heap type: dynamically created per interpreter
+            collections.deque,        # heap type: dynamically created per interpreter
         ]
-        assert objects == expected, f"Expected {{expected!r}}, got {{objects!r}}."
+        # Check that we have the expected objects. Avoid IndexError by checking lengths first.
+        assert len(objects) == len(expected), (
+            f"Expected {{expected!r}} ({{len(expected)}}), got {{objects!r}} ({{len(objects)}})."
+        )
+        # The first ones are static types shared between interpreters.
+        assert objects[:-2] == expected[:-2], (
+            f"Expected static objects {{expected!r}} ({{len(expected)}}), got {{objects!r}} ({{len(objects)}})."
+        )
+        # The last two are heap types created per-interpreter.
+        # The expected objects are dynamically imported from `collections`.
+        assert objects[-2:] == expected[-2:], (
+            f"Expected heap objects {{expected!r}} ({{len(expected)}}), got {{objects!r}} ({{len(objects)}})."
+        )
 
         assert hasattr(m, 'MyClass'), "Module missing MyClass"
         assert hasattr(m, 'MyGlobalError'), "Module missing MyGlobalError"
@@ -240,11 +252,6 @@
 ).lstrip()
 
 
-@pytest.mark.xfail(
-    reason="Duplicate C++ type registration under multiple-interpreters, needs investigation.",
-    # raises=interpreters.ExecutionFailed,  # need to import the module
-    strict=False,
-)
 @pytest.mark.skipif(
     sys.platform.startswith("emscripten"), reason="Requires loadable modules"
 )
@@ -281,18 +288,13 @@
         ) from ex
 
 
-@pytest.mark.xfail(
-    reason="Duplicate C++ type registration under multiple-interpreters, needs investigation.",
-    raises=RuntimeError,
-    strict=False,
-)
 @pytest.mark.skipif(
     sys.platform.startswith("emscripten"), reason="Requires loadable modules"
 )
 @pytest.mark.skipif(not CONCURRENT_INTERPRETERS_SUPPORT, reason="Requires 3.14.0b3+")
 def test_import_in_subinterpreter_after_main():
     """Tests that importing a module in a subinterpreter after the main interpreter works correctly"""
     check_script_success_in_subprocess(
         PREAMBLE_CODE
         + textwrap.dedent(
             """
@@ -342,18 +344,13 @@
     )
 
 
-@pytest.mark.xfail(
-    reason="Duplicate C++ type registration under multiple-interpreters, needs investigation.",
-    raises=RuntimeError,
-    strict=False,
-)
 @pytest.mark.skipif(
     sys.platform.startswith("emscripten"), reason="Requires loadable modules"
 )
 @pytest.mark.skipif(not CONCURRENT_INTERPRETERS_SUPPORT, reason="Requires 3.14.0b3+")
 def test_import_in_subinterpreter_before_main():
     """Tests that importing a module in a subinterpreter before the main interpreter works correctly"""
     check_script_success_in_subprocess(
         PREAMBLE_CODE
         + textwrap.dedent(
             """
@@ -427,18 +424,13 @@
     )
 
 
-@pytest.mark.xfail(
-    reason="Duplicate C++ type registration under multiple-interpreters, needs investigation.",
-    raises=RuntimeError,
-    strict=False,
-)
 @pytest.mark.skipif(
     sys.platform.startswith("emscripten"), reason="Requires loadable modules"
 )
 @pytest.mark.skipif(not CONCURRENT_INTERPRETERS_SUPPORT, reason="Requires 3.14.0b3+")
 def test_import_in_subinterpreter_concurrently():
     """Tests that importing a module in multiple subinterpreters concurrently works correctly"""
     check_script_success_in_subprocess(
         PREAMBLE_CODE
         + textwrap.dedent(
             """

tests/test_with_catch/test_subinterpreter.cpp

@@ -31,7 +31,7 @@ void unsafe_reset_internals_for_single_interpreter() {
     py::detail::get_local_internals_pp_manager().unref();
 
     // we know there are no other interpreters, so we can lower this. SUPER DANGEROUS
-    py::detail::get_num_interpreters_seen() = 1;
+    py::detail::get_multiple_interpreters_seen() = false;
 
     // now we unref the static global singleton internals
     py::detail::get_internals_pp_manager().unref();