fix: support free-threaded CPython with GIL disabled (#5148)

* Support free-threaded CPython (PEP 703)

Some additional locking is added in the free-threaded build when
`Py_GIL_DISABLED` is defined:

- Most accesses to internals are protected by a single mutex
- The registered_instances uses a striped lock to improve concurrency

Pybind11 modules can indicate support for running with the GIL disabled
by calling `set_gil_not_used()`.

* refactor: use PYBIND11_MODULE (#11)

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>

* Address code review

* Suppress MSVC warning

* Changes from review

* style: pre-commit fixes

* `py::mod_gil_not_used()` suggestion.

* Update include/pybind11/pybind11.h

---------

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>
Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rwgk@google.com>

include/pybind11/detail/class.h

@@ -205,8 +205,8 @@ extern "C" inline PyObject *pybind11_meta_call(PyObject *type, PyObject *args, P
 
 /// Cleanup the type-info for a pybind11-registered type.
 extern "C" inline void pybind11_meta_dealloc(PyObject *obj) {
+    with_internals([obj](internals &internals) {
         auto *type = (PyTypeObject *) obj;
-    auto &internals = get_internals();
 
         // A pybind11-registered type will:
         // 1) be found in internals.registered_types_py
@@ -238,6 +238,7 @@ extern "C" inline void pybind11_meta_dealloc(PyObject *obj) {
 
             delete tinfo;
         }
+    });
 
     PyType_Type.tp_dealloc(obj);
 }
@@ -310,19 +311,20 @@ inline void traverse_offset_bases(void *valueptr,
 }
 
 inline bool register_instance_impl(void *ptr, instance *self) {
-    get_internals().registered_instances.emplace(ptr, self);
+    with_instance_map(ptr, [&](instance_map &instances) { instances.emplace(ptr, self); });
     return true; // unused, but gives the same signature as the deregister func
 }
 inline bool deregister_instance_impl(void *ptr, instance *self) {
-    auto &registered_instances = get_internals().registered_instances;
-    auto range = registered_instances.equal_range(ptr);
+    return with_instance_map(ptr, [&](instance_map &instances) {
+        auto range = instances.equal_range(ptr);
         for (auto it = range.first; it != range.second; ++it) {
             if (self == it->second) {
-            registered_instances.erase(it);
+                instances.erase(it);
                 return true;
             }
         }
         return false;
+    });
 }
 
 inline void register_instance(instance *self, void *valptr, const type_info *tinfo) {
@@ -377,27 +379,32 @@ extern "C" inline int pybind11_object_init(PyObject *self, PyObject *, PyObject
 }
 
 inline void add_patient(PyObject *nurse, PyObject *patient) {
-    auto &internals = get_internals();
     auto *instance = reinterpret_cast<detail::instance *>(nurse);
     instance->has_patients = true;
     Py_INCREF(patient);
-    internals.patients[nurse].push_back(patient);
+
+    with_internals([&](internals &internals) { internals.patients[nurse].push_back(patient); });
 }
 
 inline void clear_patients(PyObject *self) {
     auto *instance = reinterpret_cast<detail::instance *>(self);
-    auto &internals = get_internals();
+    std::vector<PyObject *> patients;
+
+    with_internals([&](internals &internals) {
         auto pos = internals.patients.find(self);
 
         if (pos == internals.patients.end()) {
-        pybind11_fail("FATAL: Internal consistency check failed: Invalid clear_patients() call.");
+            pybind11_fail(
+                "FATAL: Internal consistency check failed: Invalid clear_patients() call.");
         }
 
         // Clearing the patients can cause more Python code to run, which
         // can invalidate the iterator. Extract the vector of patients
         // from the unordered_map first.
-    auto patients = std::move(pos->second);
+        patients = std::move(pos->second);
         internals.patients.erase(pos);
+    });
+
     instance->has_patients = false;
     for (PyObject *&patient : patients) {
         Py_CLEAR(patient);
@@ -662,10 +669,13 @@ inline PyObject *make_new_python_type(const type_record &rec) {
 
     char *tp_doc = nullptr;
     if (rec.doc && options::show_user_defined_docstrings()) {
-        /* Allocate memory for docstring (using PyObject_MALLOC, since
-           Python will free this later on) */
+        /* Allocate memory for docstring (Python will free this later on) */
         size_t size = std::strlen(rec.doc) + 1;
+#if PY_VERSION_HEX >= 0x030D0000
+        tp_doc = (char *) PyMem_MALLOC(size);
+#else
         tp_doc = (char *) PyObject_MALLOC(size);
+#endif
         std::memcpy((void *) tp_doc, rec.doc, size);
     }
 

include/pybind11/detail/common.h

@@ -464,7 +464,7 @@ PYBIND11_WARNING_POP
             });
         }
 \endrst */
-#define PYBIND11_MODULE(name, variable)                                                           \
+#define PYBIND11_MODULE(name, variable, ...)                                                      \
     static ::pybind11::module_::module_def PYBIND11_CONCAT(pybind11_module_def_, name)            \
         PYBIND11_MAYBE_UNUSED;                                                                    \
     PYBIND11_MAYBE_UNUSED                                                                         \
@@ -473,7 +473,10 @@ PYBIND11_WARNING_POP
         PYBIND11_CHECK_PYTHON_VERSION                                                             \
         PYBIND11_ENSURE_INTERNALS_READY                                                           \
         auto m = ::pybind11::module_::create_extension_module(                                    \
-            PYBIND11_TOSTRING(name), nullptr, &PYBIND11_CONCAT(pybind11_module_def_, name));      \
+            PYBIND11_TOSTRING(name),                                                              \
+            nullptr,                                                                              \
+            &PYBIND11_CONCAT(pybind11_module_def_, name),                                         \
+            ##__VA_ARGS__);                                                                       \
         try {                                                                                     \
             PYBIND11_CONCAT(pybind11_init_, name)(m);                                             \
             return m.ptr();                                                                       \

include/pybind11/detail/internals.h

@@ -18,6 +18,8 @@
 #include "../pytypes.h"
 
 #include <exception>
+#include <mutex>
+#include <thread>
 
 /// Tracks the `internals` and `type_info` ABI version independent of the main library version.
 ///
@@ -168,15 +170,37 @@ struct override_hash {
     }
 };
 
+using instance_map = std::unordered_multimap<const void *, instance *>;
+
+// ignore: structure was padded due to alignment specifier
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_MSVC(4324)
+
+// Instance map shards are used to reduce mutex contention in free-threaded Python.
+struct alignas(64) instance_map_shard {
+    std::mutex mutex;
+    instance_map registered_instances;
+};
+
+PYBIND11_WARNING_POP
+
 /// Internal data structure used to track registered instances and types.
 /// Whenever binary incompatible changes are made to this structure,
 /// `PYBIND11_INTERNALS_VERSION` must be incremented.
 struct internals {
+#ifdef Py_GIL_DISABLED
+    std::mutex mutex;
+#endif
     // std::type_index -> pybind11's type information
     type_map<type_info *> registered_types_cpp;
     // PyTypeObject* -> base type_info(s)
     std::unordered_map<PyTypeObject *, std::vector<type_info *>> registered_types_py;
-    std::unordered_multimap<const void *, instance *> registered_instances; // void * -> instance*
+#ifdef Py_GIL_DISABLED
+    std::unique_ptr<instance_map_shard[]> instance_shards; // void * -> instance*
+    size_t instance_shards_mask;
+#else
+    instance_map registered_instances; // void * -> instance*
+#endif
     std::unordered_set<std::pair<const PyObject *, const char *>, override_hash>
         inactive_override_cache;
     type_map<std::vector<bool (*)(PyObject *, void *&)>> direct_conversions;
@@ -462,7 +486,8 @@ inline object get_python_state_dict() {
 }
 
 inline object get_internals_obj_from_state_dict(handle state_dict) {
-    return reinterpret_borrow<object>(dict_getitemstring(state_dict.ptr(), PYBIND11_INTERNALS_ID));
+    return reinterpret_steal<object>(
+        dict_getitemstringref(state_dict.ptr(), PYBIND11_INTERNALS_ID));
 }
 
 inline internals **get_internals_pp_from_capsule(handle obj) {
@@ -474,6 +499,20 @@ inline internals **get_internals_pp_from_capsule(handle obj) {
     return static_cast<internals **>(raw_ptr);
 }
 
+inline uint64_t round_up_to_next_pow2(uint64_t x) {
+    // Round-up to the next power of two.
+    // See https://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2
+    x--;
+    x |= (x >> 1);
+    x |= (x >> 2);
+    x |= (x >> 4);
+    x |= (x >> 8);
+    x |= (x >> 16);
+    x |= (x >> 32);
+    x++;
+    return x;
+}
+
 /// Return a reference to the current `internals` data
 PYBIND11_NOINLINE internals &get_internals() {
     auto **&internals_pp = get_internals_pp();
@@ -542,6 +581,16 @@ PYBIND11_NOINLINE internals &get_internals() {
         internals_ptr->static_property_type = make_static_property_type();
         internals_ptr->default_metaclass = make_default_metaclass();
         internals_ptr->instance_base = make_object_base_type(internals_ptr->default_metaclass);
+#ifdef Py_GIL_DISABLED
+        // Scale proportional to the number of cores. 2x is a heuristic to reduce contention.
+        auto num_shards
+            = static_cast<size_t>(round_up_to_next_pow2(2 * std::thread::hardware_concurrency()));
+        if (num_shards == 0) {
+            num_shards = 1;
+        }
+        internals_ptr->instance_shards.reset(new instance_map_shard[num_shards]);
+        internals_ptr->instance_shards_mask = num_shards - 1;
+#endif // Py_GIL_DISABLED
     }
     return **internals_pp;
 }
@@ -602,13 +651,80 @@ inline local_internals &get_local_internals() {
     return *locals;
 }
 
+#ifdef Py_GIL_DISABLED
+#    define PYBIND11_LOCK_INTERNALS(internals) std::unique_lock<std::mutex> lock((internals).mutex)
+#else
+#    define PYBIND11_LOCK_INTERNALS(internals)
+#endif
+
+template <typename F>
+inline auto with_internals(const F &cb) -> decltype(cb(get_internals())) {
+    auto &internals = get_internals();
+    PYBIND11_LOCK_INTERNALS(internals);
+    return cb(internals);
+}
+
+inline std::uint64_t mix64(std::uint64_t z) {
+    // David Stafford's variant 13 of the MurmurHash3 finalizer popularized
+    // by the SplitMix PRNG.
+    // https://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html
+    z = (z ^ (z >> 30)) * 0xbf58476d1ce4e5b9;
+    z = (z ^ (z >> 27)) * 0x94d049bb133111eb;
+    return z ^ (z >> 31);
+}
+
+template <typename F>
+inline auto with_instance_map(const void *ptr,
+                              const F &cb) -> decltype(cb(std::declval<instance_map &>())) {
+    auto &internals = get_internals();
+
+#ifdef Py_GIL_DISABLED
+    // Hash address to compute shard, but ignore low bits. We'd like allocations
+    // from the same thread/core to map to the same shard and allocations from
+    // other threads/cores to map to other shards. Using the high bits is a good
+    // heuristic because memory allocators often have a per-thread
+    // arena/superblock/segment from which smaller allocations are served.
+    auto addr = reinterpret_cast<std::uintptr_t>(ptr);
+    auto hash = mix64(static_cast<std::uint64_t>(addr >> 20));
+    auto idx = static_cast<size_t>(hash & internals.instance_shards_mask);
+
+    auto &shard = internals.instance_shards[idx];
+    std::unique_lock<std::mutex> lock(shard.mutex);
+    return cb(shard.registered_instances);
+#else
+    (void) ptr;
+    return cb(internals.registered_instances);
+#endif
+}
+
+// Returns the number of registered instances for testing purposes.  The result may not be
+// consistent if other threads are registering or unregistering instances concurrently.
+inline size_t num_registered_instances() {
+    auto &internals = get_internals();
+#ifdef Py_GIL_DISABLED
+    size_t count = 0;
+    for (size_t i = 0; i <= internals.instance_shards_mask; ++i) {
+        auto &shard = internals.instance_shards[i];
+        std::unique_lock<std::mutex> lock(shard.mutex);
+        count += shard.registered_instances.size();
+    }
+    return count;
+#else
+    return internals.registered_instances.size();
+#endif
+}
+
 /// Constructs a std::string with the given arguments, stores it in `internals`, and returns its
 /// `c_str()`.  Such strings objects have a long storage duration -- the internal strings are only
 /// cleared when the program exits or after interpreter shutdown (when embedding), and so are
 /// suitable for c-style strings needed by Python internals (such as PyTypeObject's tp_name).
 template <typename... Args>
 const char *c_str(Args &&...args) {
-    auto &strings = get_internals().static_strings;
+    // GCC 4.8 doesn't like parameter unpack within lambda capture, so use
+    // PYBIND11_LOCK_INTERNALS.
+    auto &internals = get_internals();
+    PYBIND11_LOCK_INTERNALS(internals);
+    auto &strings = internals.static_strings;
     strings.emplace_front(std::forward<Args>(args)...);
     return strings.front().c_str();
 }
@@ -638,15 +754,18 @@ PYBIND11_NAMESPACE_END(detail)
 /// pybind11 version) running in the current interpreter. Names starting with underscores
 /// are reserved for internal usage. Returns `nullptr` if no matching entry was found.
 PYBIND11_NOINLINE void *get_shared_data(const std::string &name) {
-    auto &internals = detail::get_internals();
+    return detail::with_internals([&](detail::internals &internals) {
         auto it = internals.shared_data.find(name);
         return it != internals.shared_data.end() ? it->second : nullptr;
+    });
 }
 
 /// Set the shared data that can be later recovered by `get_shared_data()`.
 PYBIND11_NOINLINE void *set_shared_data(const std::string &name, void *data) {
-    detail::get_internals().shared_data[name] = data;
+    return detail::with_internals([&](detail::internals &internals) {
+        internals.shared_data[name] = data;
         return data;
+    });
 }
 
 /// Returns a typed reference to a shared data entry (by using `get_shared_data()`) if
@@ -654,14 +773,15 @@ PYBIND11_NOINLINE void *set_shared_data(const std::string &name, void *data) {
 /// added to the shared data under the given name and a reference to it is returned.
 template <typename T>
 T &get_or_create_shared_data(const std::string &name) {
-    auto &internals = detail::get_internals();
+    return *detail::with_internals([&](detail::internals &internals) {
         auto it = internals.shared_data.find(name);
         T *ptr = (T *) (it != internals.shared_data.end() ? it->second : nullptr);
         if (!ptr) {
             ptr = new T();
             internals.shared_data[name] = ptr;
         }
-    return *ptr;
+        return ptr;
+    });
 }
 
 PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

include/pybind11/detail/type_caster_base.h

@@ -217,12 +217,15 @@ inline detail::type_info *get_local_type_info(const std::type_index &tp) {
 }
 
 inline detail::type_info *get_global_type_info(const std::type_index &tp) {
-    auto &types = get_internals().registered_types_cpp;
+    return with_internals([&](internals &internals) {
+        detail::type_info *type_info = nullptr;
+        auto &types = internals.registered_types_cpp;
         auto it = types.find(tp);
         if (it != types.end()) {
-        return it->second;
+            type_info = it->second;
         }
-    return nullptr;
+        return type_info;
+    });
 }
 
 /// Return the type info for a given C++ type; on lookup failure can either throw or return
@@ -253,7 +256,8 @@ PYBIND11_NOINLINE handle get_type_handle(const std::type_info &tp, bool throw_if
 // Searches the inheritance graph for a registered Python instance, using all_type_info().
 PYBIND11_NOINLINE handle find_registered_python_instance(void *src,
                                                          const detail::type_info *tinfo) {
-    auto it_instances = get_internals().registered_instances.equal_range(src);
+    return with_instance_map(src, [&](instance_map &instances) {
+        auto it_instances = instances.equal_range(src);
         for (auto it_i = it_instances.first; it_i != it_instances.second; ++it_i) {
             for (auto *instance_type : detail::all_type_info(Py_TYPE(it_i->second))) {
                 if (instance_type && same_type(*instance_type->cpptype, *tinfo->cpptype)) {
@@ -262,6 +266,7 @@ PYBIND11_NOINLINE handle find_registered_python_instance(void *src,
             }
         }
         return handle();
+    });
 }
 
 struct value_and_holder {
@@ -506,7 +511,7 @@ PYBIND11_NOINLINE bool isinstance_generic(handle obj, const std::type_info &tp)
 }
 
 PYBIND11_NOINLINE handle get_object_handle(const void *ptr, const detail::type_info *type) {
-    auto &instances = get_internals().registered_instances;
+    return with_instance_map(ptr, [&](instance_map &instances) {
         auto range = instances.equal_range(ptr);
         for (auto it = range.first; it != range.second; ++it) {
             for (const auto &vh : values_and_holders(it->second)) {
@@ -516,6 +521,7 @@ PYBIND11_NOINLINE handle get_object_handle(const void *ptr, const detail::type_i
             }
         }
         return handle();
+    });
 }
 
 inline PyThreadState *get_thread_state_unchecked() {

include/pybind11/numpy.h

@@ -1553,7 +1553,9 @@ PYBIND11_NOINLINE void register_structured_dtype(any_container<field_descriptor>
 
     auto tindex = std::type_index(tinfo);
     numpy_internals.registered_dtypes[tindex] = {dtype_ptr, std::move(format_str)};
-    get_internals().direct_conversions[tindex].push_back(direct_converter);
+    with_internals([tindex, &direct_converter](internals &internals) {
+        internals.direct_conversions[tindex].push_back(direct_converter);
+    });
 }
 
 template <typename T, typename SFINAE>

include/pybind11/pybind11.h

@@ -1054,13 +1054,20 @@ protected:
                 - delegate translation to the next translator by throwing a new type of exception.
              */
 
+            bool handled = with_internals([&](internals &internals) {
                 auto &local_exception_translators
                     = get_local_internals().registered_exception_translators;
                 if (detail::apply_exception_translators(local_exception_translators)) {
-                return nullptr;
+                    return true;
                 }
-            auto &exception_translators = get_internals().registered_exception_translators;
+                auto &exception_translators = internals.registered_exception_translators;
                 if (detail::apply_exception_translators(exception_translators)) {
+                    return true;
+                }
+                return false;
+            });
+
+            if (handled) {
                 return nullptr;
             }
 
@@ -1199,6 +1206,16 @@ struct handle_type_name<cpp_function> {
 
 PYBIND11_NAMESPACE_END(detail)
 
+// Use to activate Py_MOD_GIL_NOT_USED.
+class mod_gil_not_used {
+public:
+    explicit mod_gil_not_used(bool flag = true) : flag_(flag) {}
+    bool flag() const { return flag_; }
+
+private:
+    bool flag_;
+};
+
 /// Wrapper for Python extension modules
 class module_ : public object {
 public:
@@ -1299,7 +1316,11 @@ public:
 
         ``def`` should point to a statically allocated module_def.
     \endrst */
-    static module_ create_extension_module(const char *name, const char *doc, module_def *def) {
+    static module_ create_extension_module(const char *name,
+                                           const char *doc,
+                                           module_def *def,
+                                           mod_gil_not_used gil_not_used
+                                           = mod_gil_not_used(false)) {
         // module_def is PyModuleDef
         // Placement new (not an allocation).
         def = new (def)
@@ -1319,6 +1340,11 @@ public:
             }
             pybind11_fail("Internal error in module_::create_extension_module()");
         }
+        if (gil_not_used.flag()) {
+#ifdef Py_GIL_DISABLED
+            PyUnstable_Module_SetGIL(m, Py_MOD_GIL_NOT_USED);
+#endif
+        }
         // TODO: Should be reinterpret_steal for Python 3, but Python also steals it again when
         //       returned from PyInit_...
         //       For Python 2, reinterpret_borrow was correct.
@@ -1397,7 +1423,7 @@ protected:
         tinfo->default_holder = rec.default_holder;
         tinfo->module_local = rec.module_local;
 
-        auto &internals = get_internals();
+        with_internals([&](internals &internals) {
             auto tindex = std::type_index(*rec.type);
             tinfo->direct_conversions = &internals.direct_conversions[tindex];
             if (rec.module_local) {
@@ -1406,6 +1432,7 @@ protected:
                 internals.registered_types_cpp[tindex] = tinfo;
             }
             internals.registered_types_py[(PyTypeObject *) m_ptr] = {tinfo};
+        });
 
         if (rec.bases.size() > 1 || rec.multiple_inheritance) {
             mark_parents_nonsimple(tinfo->type);
@@ -1618,10 +1645,12 @@ public:
         generic_type::initialize(record);
 
         if (has_alias) {
+            with_internals([&](internals &internals) {
                 auto &instances = record.module_local ? get_local_internals().registered_types_cpp
-                                                  : get_internals().registered_types_cpp;
+                                                      : internals.registered_types_cpp;
                 instances[std::type_index(typeid(type_alias))]
                     = instances[std::type_index(typeid(type))];
+            });
         }
     }
 
@@ -2336,21 +2365,24 @@ keep_alive_impl(size_t Nurse, size_t Patient, function_call &call, handle ret) {
 
 inline std::pair<decltype(internals::registered_types_py)::iterator, bool>
 all_type_info_get_cache(PyTypeObject *type) {
-    auto res = get_internals()
+    auto res = with_internals([type](internals &internals) {
+        return internals
             .registered_types_py
 #ifdef __cpp_lib_unordered_map_try_emplace
             .try_emplace(type);
 #else
             .emplace(type, std::vector<detail::type_info *>());
 #endif
+    });
     if (res.second) {
         // New cache entry created; set up a weak reference to automatically remove it if the type
         // gets destroyed:
         weakref((PyObject *) type, cpp_function([type](handle wr) {
-                    get_internals().registered_types_py.erase(type);
+                    with_internals([type](internals &internals) {
+                        internals.registered_types_py.erase(type);
 
                         // TODO consolidate the erasure code in pybind11_meta_dealloc() in class.h
-                    auto &cache = get_internals().inactive_override_cache;
+                        auto &cache = internals.inactive_override_cache;
                         for (auto it = cache.begin(), last = cache.end(); it != last;) {
                             if (it->first == reinterpret_cast<PyObject *>(type)) {
                                 it = cache.erase(it);
@@ -2358,6 +2390,7 @@ all_type_info_get_cache(PyTypeObject *type) {
                                 ++it;
                             }
                         }
+                    });
 
                     wr.dec_ref();
                 }))
@@ -2562,7 +2595,11 @@ void implicitly_convertible() {
         ~set_flag() { flag = false; }
     };
     auto implicit_caster = [](PyObject *obj, PyTypeObject *type) -> PyObject * {
+#ifdef Py_GIL_DISABLED
+        thread_local bool currently_used = false;
+#else
         static bool currently_used = false;
+#endif
         if (currently_used) { // implicit conversions are non-reentrant
             return nullptr;
         }
@@ -2587,8 +2624,10 @@ void implicitly_convertible() {
 }
 
 inline void register_exception_translator(ExceptionTranslator &&translator) {
-    detail::get_internals().registered_exception_translators.push_front(
+    detail::with_internals([&](detail::internals &internals) {
+        internals.registered_exception_translators.push_front(
             std::forward<ExceptionTranslator>(translator));
+    });
 }
 
 /**
@@ -2598,8 +2637,11 @@ inline void register_exception_translator(ExceptionTranslator &&translator) {
  * the exception.
  */
 inline void register_local_exception_translator(ExceptionTranslator &&translator) {
+    detail::with_internals([&](detail::internals &internals) {
+        (void) internals;
         detail::get_local_internals().registered_exception_translators.push_front(
             std::forward<ExceptionTranslator>(translator));
+    });
 }
 
 /**
@@ -2756,14 +2798,19 @@ get_type_override(const void *this_ptr, const type_info *this_type, const char *
 
     /* Cache functions that aren't overridden in Python to avoid
        many costly Python dictionary lookups below */
-    auto &cache = get_internals().inactive_override_cache;
-    if (cache.find(key) != cache.end()) {
+    bool not_overridden = with_internals([&key](internals &internals) {
+        auto &cache = internals.inactive_override_cache;
+        return cache.find(key) != cache.end();
+    });
+    if (not_overridden) {
         return function();
     }
 
     function override = getattr(self, name, function());
     if (override.is_cpp_function()) {
-        cache.insert(std::move(key));
+        with_internals([&](internals &internals) {
+            internals.inactive_override_cache.insert(std::move(key));
+        });
         return function();
     }
 

include/pybind11/pytypes.h

@@ -980,6 +980,23 @@ inline PyObject *dict_getitem(PyObject *v, PyObject *key) {
     return rv;
 }
 
+inline PyObject *dict_getitemstringref(PyObject *v, const char *key) {
+#if PY_VERSION_HEX >= 0x030D0000
+    PyObject *rv;
+    if (PyDict_GetItemStringRef(v, key, &rv) < 0) {
+        throw error_already_set();
+    }
+    return rv;
+#else
+    PyObject *rv = dict_getitemstring(v, key);
+    if (rv == nullptr && PyErr_Occurred()) {
+        throw error_already_set();
+    }
+    Py_XINCREF(rv);
+    return rv;
+#endif
+}
+
 // Helper aliases/functions to support implicit casting of values given to python
 // accessors/methods. When given a pyobject, this simply returns the pyobject as-is; for other C++
 // type, the value goes through pybind11::cast(obj) to convert it to an `object`.

tests/cross_module_gil_utils.cpp

@@ -92,6 +92,9 @@ extern "C" PYBIND11_EXPORT PyObject *PyInit_cross_module_gil_utils() {
     if (m != nullptr) {
         static_assert(sizeof(&gil_acquire) == sizeof(void *),
                       "Function pointer must have the same size as void*");
+#ifdef Py_GIL_DISABLED
+        PyUnstable_Module_SetGIL(m, Py_MOD_GIL_NOT_USED);
+#endif
         ADD_FUNCTION("gil_acquire_funcaddr", gil_acquire)
         ADD_FUNCTION("gil_multi_acquire_release_funcaddr", gil_multi_acquire_release)
         ADD_FUNCTION("gil_acquire_inner_custom_funcaddr",

tests/cross_module_interleaved_error_already_set.cpp

@@ -42,6 +42,9 @@ extern "C" PYBIND11_EXPORT PyObject *PyInit_cross_module_interleaved_error_alrea
     if (m != nullptr) {
         static_assert(sizeof(&interleaved_error_already_set) == sizeof(void *),
                       "Function pointer must have the same size as void *");
+#ifdef Py_GIL_DISABLED
+        PyUnstable_Module_SetGIL(m, Py_MOD_GIL_NOT_USED);
+#endif
         PyModule_AddObject(
             m,
             "funcaddr",

tests/eigen_tensor_avoid_stl_array.cpp

@@ -11,4 +11,6 @@
 
 #include "test_eigen_tensor.inl"
 
-PYBIND11_MODULE(eigen_tensor_avoid_stl_array, m) { eigen_tensor_test::test_module(m); }
+PYBIND11_MODULE(eigen_tensor_avoid_stl_array, m, pybind11::mod_gil_not_used()) {
+    eigen_tensor_test::test_module(m);
+}

tests/env.py

@@ -1,5 +1,6 @@
 import platform
 import sys
+import sysconfig
 
 import pytest
 
@@ -9,6 +10,7 @@ WIN = sys.platform.startswith("win32") or sys.platform.startswith("cygwin")
 
 CPYTHON = platform.python_implementation() == "CPython"
 PYPY = platform.python_implementation() == "PyPy"
+PY_GIL_DISABLED = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
 
 
 def deprecated_call():

tests/pybind11_cross_module_tests.cpp

@@ -16,7 +16,7 @@
 #include <numeric>
 #include <utility>
 
-PYBIND11_MODULE(pybind11_cross_module_tests, m) {
+PYBIND11_MODULE(pybind11_cross_module_tests, m, py::mod_gil_not_used()) {
     m.doc() = "pybind11 cross-module test module";
 
     // test_local_bindings.py tests:

tests/pybind11_tests.cpp

@@ -58,7 +58,7 @@ void bind_ConstructorStats(py::module_ &m) {
         // registered instances to allow instance cleanup checks (invokes a GC first)
         .def_static("detail_reg_inst", []() {
             ConstructorStats::gc();
-            return py::detail::get_internals().registered_instances.size();
+            return py::detail::num_registered_instances();
         });
 }
 
@@ -75,7 +75,7 @@ const char *cpp_std() {
 #endif
 }
 
-PYBIND11_MODULE(pybind11_tests, m) {
+PYBIND11_MODULE(pybind11_tests, m, py::mod_gil_not_used()) {
     m.doc() = "pybind11 test module";
 
     // Intentionally kept minimal to not create a maintenance chore

tests/test_cmake_build/main.cpp

@@ -1,6 +1,6 @@
 #include <pybind11/pybind11.h>
 namespace py = pybind11;
 
-PYBIND11_MODULE(test_cmake_build, m) {
+PYBIND11_MODULE(test_cmake_build, m, py::mod_gil_not_used()) {
     m.def("add", [](int i, int j) { return i + j; });
 }

tests/test_embed/external_module.cpp

@@ -6,7 +6,7 @@ namespace py = pybind11;
  * modules aren't preserved over a finalize/initialize.
  */
 
-PYBIND11_MODULE(external_module, m) {
+PYBIND11_MODULE(external_module, m, py::mod_gil_not_used()) {
     class A {
     public:
         explicit A(int value) : v{value} {};