include/pybind11/detail/internals.h

/*
    pybind11/detail/internals.h: Internal data structure and related functions

    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>

    All rights reserved. Use of this source code is governed by a
    BSD-style license that can be found in the LICENSE file.
*/

#pragma once

#include <pybind11/conduit/pybind11_platform_abi_id.h>
#include <pybind11/gil_simple.h>
#include <pybind11/pytypes.h>
#include <pybind11/trampoline_self_life_support.h>

#include "common.h"
#include "struct_smart_holder.h"

#include <atomic>
#include <cstdint>
#include <exception>
#include <limits>
#include <mutex>
#include <thread>

/// Tracks the `internals` and `type_info` ABI version independent of the main library version.
///
/// Some portions of the code use an ABI that is conditional depending on this
/// version number.  That allows ABI-breaking changes to be "pre-implemented".
/// Once the default version number is incremented, the conditional logic that
/// no longer applies can be removed.  Additionally, users that need not
/// maintain ABI compatibility can increase the version number in order to take
/// advantage of any functionality/efficiency improvements that depend on the
/// newer ABI.
///
/// WARNING: If you choose to manually increase the ABI version, note that
/// pybind11 may not be tested as thoroughly with a non-default ABI version, and
/// further ABI-incompatible changes may be made before the ABI is officially
/// changed to the new version.
#ifndef PYBIND11_INTERNALS_VERSION
#    define PYBIND11_INTERNALS_VERSION 12
#endif

#if PYBIND11_INTERNALS_VERSION < 12
#    error "PYBIND11_INTERNALS_VERSION 12 is the minimum for all platforms for pybind11 v3.1.0"
#endif

PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)

using ExceptionTranslator = void (*)(std::exception_ptr);

// The old Python Thread Local Storage (TLS) API is deprecated in Python 3.7 in favor of the new
// Thread Specific Storage (TSS) API.
// Avoid unnecessary allocation of `Py_tss_t`, since we cannot use
// `Py_LIMITED_API` anyway.
#define PYBIND11_TLS_KEY_REF Py_tss_t &
#if defined(__clang__)
#    define PYBIND11_TLS_KEY_INIT(var)                                                            \
        _Pragma("clang diagnostic push")                                         /**/             \
            _Pragma("clang diagnostic ignored \"-Wmissing-field-initializers\"") /**/             \
            Py_tss_t var = Py_tss_NEEDS_INIT;                                                     \
        _Pragma("clang diagnostic pop")
#elif defined(__GNUC__) && !defined(__INTEL_COMPILER)
#    define PYBIND11_TLS_KEY_INIT(var)                                                            \
        _Pragma("GCC diagnostic push")                                         /**/               \
            _Pragma("GCC diagnostic ignored \"-Wmissing-field-initializers\"") /**/               \
            Py_tss_t var = Py_tss_NEEDS_INIT;                                                     \
        _Pragma("GCC diagnostic pop")
#else
#    define PYBIND11_TLS_KEY_INIT(var) Py_tss_t var = Py_tss_NEEDS_INIT;
#endif
#define PYBIND11_TLS_KEY_CREATE(var) (PyThread_tss_create(&(var)) == 0)
#define PYBIND11_TLS_GET_VALUE(key) PyThread_tss_get(&(key))
#define PYBIND11_TLS_REPLACE_VALUE(key, value) PyThread_tss_set(&(key), (value))
#define PYBIND11_TLS_DELETE_VALUE(key) PyThread_tss_set(&(key), nullptr)
#define PYBIND11_TLS_FREE(key) PyThread_tss_delete(&(key))

/// A smart-pointer-like wrapper around a thread-specific value. get/set of the pointer applies to
/// the current thread only.
template <typename T>
class thread_specific_storage {
public:
    thread_specific_storage() {
        // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
        if (!PYBIND11_TLS_KEY_CREATE(key_)) {
            pybind11_fail(
                "thread_specific_storage constructor: could not initialize the TSS key!");
        }
    }

    ~thread_specific_storage() {
        // This destructor is often called *after* Py_Finalize(). That *SHOULD BE* fine on most
        // platforms. The following details what happens when PyThread_tss_free is called in
        // CPython. PYBIND11_TLS_FREE is PyThread_tss_free on python 3.7+. On older python, it does
        // nothing. PyThread_tss_free calls PyThread_tss_delete and PyMem_RawFree.
        // PyThread_tss_delete just calls TlsFree (on Windows) or pthread_key_delete (on *NIX).
        // Neither of those have anything to do with CPython internals. PyMem_RawFree *requires*
        // that the `key` be allocated with the CPython allocator (as it is by
        // PyThread_tss_create).
        // However, in GraalPy (as of v25.0 or older), TSS is implemented by Java and this call
        // requires a living Python interpreter.
#ifdef GRAALVM_PYTHON
        if (Py_IsInitialized() == 0 || _Py_IsFinalizing() != 0) {
            return;
        }
#endif
        PYBIND11_TLS_FREE(key_);
    }

    thread_specific_storage(thread_specific_storage const &) = delete;
    thread_specific_storage(thread_specific_storage &&) = delete;
    thread_specific_storage &operator=(thread_specific_storage const &) = delete;
    thread_specific_storage &operator=(thread_specific_storage &&) = delete;

    T *get() const { return reinterpret_cast<T *>(PYBIND11_TLS_GET_VALUE(key_)); }

    T &operator*() const { return *get(); }
    explicit operator T *() const { return get(); }
    explicit operator bool() const { return get() != nullptr; }

    void set(T *val) { PYBIND11_TLS_REPLACE_VALUE(key_, reinterpret_cast<void *>(val)); }
    void reset(T *p = nullptr) { set(p); }
    thread_specific_storage &operator=(T *pval) {
        set(pval);
        return *this;
    }

private:
    PYBIND11_TLS_KEY_INIT(mutable key_)
};

PYBIND11_NAMESPACE_BEGIN(detail)

// This does NOT actually exist as a module.
#define PYBIND11_DUMMY_MODULE_NAME "pybind11_builtins"

// Forward declarations
inline PyTypeObject *make_static_property_type();
inline PyTypeObject *make_default_metaclass();
inline PyObject *make_object_base_type(PyTypeObject *metaclass);
inline void translate_exception(std::exception_ptr p);

inline PyThreadState *get_thread_state_unchecked() {
#if defined(PYPY_VERSION) || defined(GRAALVM_PYTHON)
    return PyThreadState_GET();
#elif PY_VERSION_HEX < 0x030D0000
    return _PyThreadState_UncheckedGet();
#else
    return PyThreadState_GetUnchecked();
#endif
}

inline PyInterpreterState *get_interpreter_state_unchecked() {
    auto *tstate = get_thread_state_unchecked();
    return tstate ? tstate->interp : nullptr;
}

inline object get_python_state_dict() {
    object state_dict;
#if defined(PYPY_VERSION) || defined(GRAALVM_PYTHON)
    state_dict = reinterpret_borrow<object>(PyEval_GetBuiltins());
#else
    auto *istate = get_interpreter_state_unchecked();
    if (istate) {
        state_dict = reinterpret_borrow<object>(PyInterpreterState_GetDict(istate));
    }
#endif
    if (!state_dict) {
        raise_from(PyExc_SystemError, "pybind11::detail::get_python_state_dict() FAILED");
        throw error_already_set();
    }
    return state_dict;
}

// Python loads modules by default with dlopen with the RTLD_LOCAL flag; under libc++ and possibly
// other STLs, this means `typeid(A)` from one module won't equal `typeid(A)` from another module
// even when `A` is the same, non-hidden-visibility type (e.g. from a common include).  Under
// libstdc++, this doesn't happen: equality and the type_index hash are based on the type name,
// which works.  If not under a known-good stl, provide our own name-based hash and equality
// functions that use the type name.
#if !defined(_LIBCPP_VERSION)
inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) { return lhs == rhs; }
using type_hash = std::hash<std::type_index>;
using type_equal_to = std::equal_to<std::type_index>;
#else
inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) {
    return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
}

struct type_hash {
    size_t operator()(const std::type_index &t) const noexcept {
        size_t hash = 5381;
        const char *ptr = t.name();
        while (auto c = static_cast<unsigned char>(*ptr++)) {
            hash = (hash * 33) ^ c;
        }
        return hash;
    }
};

struct type_equal_to {
    bool operator()(const std::type_index &lhs, const std::type_index &rhs) const noexcept {
        return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
    }
};
#endif

// For now, we don't bother adding a fancy hash for pointers and just
// let the standard library use the identity hash function if that's
// what it wants to do (e.g., as in libstdc++).
template <typename value_type>
using fast_type_map = std::unordered_map<const std::type_info *, value_type>;

template <typename value_type>
using type_map = std::unordered_map<std::type_index, value_type, type_hash, type_equal_to>;

struct override_hash {
    size_t operator()(const std::pair<const PyObject *, const char *> &v) const noexcept {
        size_t value = std::hash<const void *>()(v.first);
        value ^= std::hash<const void *>()(v.second) + 0x9e3779b9 + (value << 6) + (value >> 2);
        return value;
    }
};

using instance_map = std::unordered_multimap<const void *, instance *>;

#ifdef Py_GIL_DISABLED
// Wrapper around PyMutex to provide BasicLockable semantics
class pymutex {
    friend class pycritical_section;
    PyMutex mutex;

public:
    pymutex() : mutex({}) {}
    void lock() { PyMutex_Lock(&mutex); }
    void unlock() { PyMutex_Unlock(&mutex); }
};

class pycritical_section {
    pymutex &mutex;
#    if PY_VERSION_HEX >= 0x030E00C1 // 3.14.0rc1
    PyCriticalSection cs;
#    endif

public:
    explicit pycritical_section(pymutex &m) : mutex(m) {
        // PyCriticalSection_BeginMutex was added in Python 3.15.0a1 and backported to 3.14.0rc1
#    if PY_VERSION_HEX >= 0x030E00C1 // 3.14.0rc1
        PyCriticalSection_BeginMutex(&cs, &mutex.mutex);
#    else
        // Fall back to direct mutex locking for older free-threaded Python versions
        mutex.lock();
#    endif
    }
    ~pycritical_section() {
#    if PY_VERSION_HEX >= 0x030E00C1 // 3.14.0rc1
        PyCriticalSection_End(&cs);
#    else
        mutex.unlock();
#    endif
    }

    // Non-copyable and non-movable to prevent double-unlock
    pycritical_section(const pycritical_section &) = delete;
    pycritical_section &operator=(const pycritical_section &) = delete;
    pycritical_section(pycritical_section &&) = delete;
    pycritical_section &operator=(pycritical_section &&) = delete;
};

// Instance map shards are used to reduce mutex contention in free-threaded Python.
struct instance_map_shard {
    instance_map registered_instances;
    pymutex mutex;
    // alignas(64) would be better, but causes compile errors in macOS before 10.14 (see #5200)
    char padding[64 - (sizeof(instance_map) + sizeof(pymutex)) % 64];
};

static_assert(sizeof(instance_map_shard) % 64 == 0,
              "instance_map_shard size is not a multiple of 64 bytes");

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;
}
#endif

class loader_life_support;

/// 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
    pymutex mutex;
    pymutex exception_translator_mutex;
#endif
#if PYBIND11_INTERNALS_VERSION >= 12
    // non-normative but fast "hint" for registered_types_cpp. Meant
    // to be used as the first level of a two-level lookup: successful
    // lookups are correct, but unsuccessful lookups need to try
    // registered_types_cpp and then backfill this map if they find
    // anything.
    fast_type_map<type_info *> registered_types_cpp_fast;
#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;
#ifdef Py_GIL_DISABLED
    std::unique_ptr<instance_map_shard[]> instance_shards; // void * -> instance*
    size_t instance_shards_mask = 0;
#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;
    std::unordered_map<const PyObject *, std::vector<PyObject *>> patients;
    std::forward_list<ExceptionTranslator> registered_exception_translators;
    std::unordered_map<std::string, void *> shared_data; // Custom data to be shared across
                                                         // extensions
    std::forward_list<std::string> static_strings;       // Stores the std::strings backing
                                                         // detail::c_str()
    PyTypeObject *static_property_type = nullptr;
    PyTypeObject *default_metaclass = nullptr;
    PyObject *instance_base = nullptr;
    // Unused if PYBIND11_SIMPLE_GIL_MANAGEMENT is defined:
    thread_specific_storage<PyThreadState> tstate;
#if PYBIND11_INTERNALS_VERSION <= 11
    thread_specific_storage<loader_life_support> loader_life_support_tls; // OBSOLETE (PR #5830)
#endif
    // Unused if PYBIND11_SIMPLE_GIL_MANAGEMENT is defined:
    PyInterpreterState *istate = nullptr;

    type_map<PyObject *> native_enum_type_map;

    internals()
        : static_property_type(make_static_property_type()),
          default_metaclass(make_default_metaclass()), istate(get_interpreter_state_unchecked()) {
        tstate.set(nullptr); // See PR #5870
        registered_exception_translators.push_front(&translate_exception);
#ifdef Py_GIL_DISABLED
        // Scale proportional to the number of cores. 2x is a heuristic to reduce contention.
        // Make sure the number isn't unreasonable by limiting it to 16 bits (65K)
        auto num_shards = static_cast<std::uint16_t>(
            std::min<std::size_t>(round_up_to_next_pow2(2 * std::thread::hardware_concurrency()),
                                  std::numeric_limits<std::uint16_t>::max()));
        if (num_shards == 0) {
            num_shards = 1;
        }
        instance_shards.reset(new instance_map_shard[num_shards]);
        instance_shards_mask = num_shards - 1;
#endif
    }
    internals(const internals &other) = delete;
    internals(internals &&other) = delete;
    internals &operator=(const internals &other) = delete;
    internals &operator=(internals &&other) = delete;
    ~internals() = default;
};

// the internals struct (above) is shared between all the modules. local_internals are only
// for a single module. Any changes made to internals may require an update to
// PYBIND11_INTERNALS_VERSION, breaking backwards compatibility. local_internals is, by design,
// restricted to a single module. Whether a module has local internals or not should not
// impact any other modules, because the only things accessing the local internals is the
// module that contains them.
struct local_internals {
    // It should be safe to use fast_type_map here because this entire
    // data structure is scoped to our single module, and thus a single
    // DSO and single instance of type_info for any particular type.
    fast_type_map<type_info *> registered_types_cpp;

    std::forward_list<ExceptionTranslator> registered_exception_translators;
    PyTypeObject *function_record_py_type = nullptr;
};

enum class holder_enum_t : uint8_t {
    undefined,
    std_unique_ptr, // Default, lacking interop with std::shared_ptr.
    std_shared_ptr, // Lacking interop with std::unique_ptr.
    smart_holder,   // Full std::unique_ptr / std::shared_ptr interop.
    custom_holder,
};

/// Additional type information which does not fit into the PyTypeObject.
/// Changes to this struct also require bumping `PYBIND11_INTERNALS_VERSION`.
struct type_info {
    PyTypeObject *type;
    const std::type_info *cpptype;
    size_t type_size, type_align, holder_size_in_ptrs;
    void *(*operator_new)(size_t);
    void (*init_instance)(instance *, const void *);
    void (*dealloc)(value_and_holder &v_h);

    // Cross-DSO-safe function pointers, to sidestep cross-DSO RTTI issues
    // on platforms like macOS (see PR #5728 for details):
    memory::get_guarded_delete_fn get_memory_guarded_delete = memory::get_guarded_delete;
    get_trampoline_self_life_support_fn get_trampoline_self_life_support = nullptr;

    std::vector<PyObject *(*) (PyObject *, PyTypeObject *)> implicit_conversions;
    std::vector<std::pair<const std::type_info *, void *(*) (void *)>> implicit_casts;
    std::vector<bool (*)(PyObject *, void *&)> *direct_conversions;
    buffer_info *(*get_buffer)(PyObject *, void *) = nullptr;
    void *get_buffer_data = nullptr;
    void *(*module_local_load)(PyObject *, const type_info *) = nullptr;
    holder_enum_t holder_enum_v = holder_enum_t::undefined;

#if PYBIND11_INTERNALS_VERSION >= 12
    // When a type appears in multiple DSOs,
    // internals::registered_types_cpp_fast will have multiple distinct
    // keys (the std::type_info from each DSO) mapped to the same
    // detail::type_info*. We need to keep track of these aliases so that we clean
    // them up when our type is deallocated. A linked list is appropriate
    // because it is expected to be 1) usually empty and 2)
    // when it's not empty, usually very small. See also `struct
    // nb_alias_chain` added in
    // https://github.com/wjakob/nanobind/commit/b515b1f7f2f4ecc0357818e6201c94a9f4cbfdc2
    std::forward_list<const std::type_info *> alias_chain;
#endif

    /* A simple type never occurs as a (direct or indirect) parent
     * of a class that makes use of multiple inheritance.
     * A type can be simple even if it has non-simple ancestors as long as it has no descendants.
     */
    bool simple_type : 1;
    /* True if there is no multiple inheritance in this type's inheritance tree */
    bool simple_ancestors : 1;
    /* true if this is a type registered with py::module_local */
    bool module_local : 1;
};

/// Information stored in a capsule on py::native_enum() types. Since we don't
/// create a type_info record for native enums, we must store here any
/// information we will need about the enum at runtime.
///
/// If you make backward-incompatible changes to this structure, you must
/// change the `attribute_name()` so that native enums from older version of
/// pybind11 don't have their records reinterpreted. Better would be to keep
/// the changes backward-compatible (i.e., only add new fields at the end)
/// and detect/indicate their presence using the currently-unused `version`.
struct native_enum_record {
    const std::type_info *cpptype;
    uint32_t size_bytes;
    bool is_signed;
    const uint8_t version = 1;

    static const char *attribute_name() { return "__pybind11_native_enum__"; }
};

#define PYBIND11_INTERNALS_ID                                                                     \
    "__pybind11_internals_v" PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION)                        \
        PYBIND11_COMPILER_TYPE_LEADING_UNDERSCORE PYBIND11_PLATFORM_ABI_ID "__"

#define PYBIND11_MODULE_LOCAL_ID                                                                  \
    "__pybind11_module_local_v" PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION)                     \
        PYBIND11_COMPILER_TYPE_LEADING_UNDERSCORE PYBIND11_PLATFORM_ABI_ID "__"

/// 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 &has_seen_non_main_interpreter() {
    static std::atomic_bool multi(false);
    return multi;
}

template <class T,
          enable_if_t<std::is_same<std::nested_exception, remove_cvref_t<T>>::value, int> = 0>
bool handle_nested_exception(const T &exc, const std::exception_ptr &p) {
    std::exception_ptr nested = exc.nested_ptr();
    if (nested != nullptr && nested != p) {
        translate_exception(nested);
        return true;
    }
    return false;
}

template <class T,
          enable_if_t<!std::is_same<std::nested_exception, remove_cvref_t<T>>::value, int> = 0>
bool handle_nested_exception(const T &exc, const std::exception_ptr &p) {
    if (const auto *nep = dynamic_cast<const std::nested_exception *>(std::addressof(exc))) {
        return handle_nested_exception(*nep, p);
    }
    return false;
}

inline bool raise_err(PyObject *exc_type, const char *msg) {
    if (PyErr_Occurred()) {
        raise_from(exc_type, msg);
        return true;
    }
    set_error(exc_type, msg);
    return false;
}

inline void translate_exception(std::exception_ptr p) {
    if (!p) {
        return;
    }
    try {
        std::rethrow_exception(p);
    } catch (error_already_set &e) {
        handle_nested_exception(e, p);
        e.restore();
        return;
    } catch (const builtin_exception &e) {
        // Could not use template since it's an abstract class.
        if (const auto *nep = dynamic_cast<const std::nested_exception *>(std::addressof(e))) {
            handle_nested_exception(*nep, p);
        }
        e.set_error();
        return;
    } catch (const std::bad_alloc &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_MemoryError, e.what());
        return;
    } catch (const std::domain_error &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_ValueError, e.what());
        return;
    } catch (const std::invalid_argument &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_ValueError, e.what());
        return;
    } catch (const std::length_error &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_ValueError, e.what());
        return;
    } catch (const std::out_of_range &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_IndexError, e.what());
        return;
    } catch (const std::range_error &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_ValueError, e.what());
        return;
    } catch (const std::overflow_error &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_OverflowError, e.what());
        return;
    } catch (const std::exception &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_RuntimeError, e.what());
        return;
    } catch (const std::nested_exception &e) {
        handle_nested_exception(e, p);
        raise_err(PyExc_RuntimeError, "Caught an unknown nested exception!");
        return;
    } catch (...) {
        raise_err(PyExc_RuntimeError, "Caught an unknown exception!");
        return;
    }
}

#if !defined(__GLIBCXX__)
inline void translate_local_exception(std::exception_ptr p) {
    try {
        if (p) {
            std::rethrow_exception(p);
        }
    } catch (error_already_set &e) {
        e.restore();
        return;
    } catch (const builtin_exception &e) {
        e.set_error();
        return;
    }
}
#endif

// Sentinel value for the `dtor` parameter of `atomic_get_or_create_in_state_dict`.
// Indicates no destructor was explicitly provided (distinct from nullptr, which means "leak").
#define PYBIND11_DTOR_USE_DELETE (reinterpret_cast<void (*)(PyObject *)>(1))

// Get or create per-storage capsule in the current interpreter's state dict.
//   - The storage is interpreter-dependent: different interpreters will have different storage.
//     This is important when using multiple-interpreters, to avoid sharing unshareable objects
//     between interpreters.
//   - There is one storage per `key` in an interpreter and it is accessible between all extensions
//     in the same interpreter.
//   - The life span of the storage is tied to the interpreter: it will be kept alive until the
//     interpreter shuts down.
//
// Use test-and-set pattern with `PyDict_SetDefault` for thread-safe concurrent access.
// WARNING: There can be multiple threads creating the storage at the same time, while only one
//          will succeed in inserting its capsule into the dict. Therefore, the deleter will be
//          used to clean up the storage of the unused capsules.
//
// Returns: pair of (pointer to storage, bool indicating if newly created).
//          The bool follows std::map::insert convention: true = created, false = existed.
// `dtor`: optional destructor called when the interpreter shuts down.
//   - If not provided: the storage will be deleted using `delete`.
//   - If nullptr: the storage will be leaked (useful for singletons that outlive the interpreter).
//   - If a function: that function will be called with the capsule object.
template <typename Payload>
std::pair<Payload *, bool> atomic_get_or_create_in_state_dict(const char *key,
                                                              void (*dtor)(PyObject *)
                                                              = PYBIND11_DTOR_USE_DELETE) {
    error_scope err_scope; // preserve any existing Python error states

    auto state_dict = reinterpret_borrow<dict>(get_python_state_dict());
    PyObject *capsule_obj = nullptr;
    bool created = false;

    // Try to get existing storage (fast path).
    capsule_obj = dict_getitemstring(state_dict.ptr(), key);
    if (capsule_obj == nullptr) {
        if (PyErr_Occurred()) {
            throw error_already_set();
        }
        // Storage doesn't exist yet, create a new one.
        // Use unique_ptr for exception safety: if capsule creation throws, the storage is
        // automatically deleted.
        auto storage_ptr = std::unique_ptr<Payload>(new Payload{});
        auto new_capsule
            = capsule(storage_ptr.get(),
                      // The destructor will be called when the capsule is GC'ed.
                      //  If the insert below fails (entry already in the dict), then this
                      //  destructor will be called on the newly created capsule at the end of this
                      //  function, and we want to just release this memory.
                      /*destructor=*/[](void *v) { delete static_cast<Payload *>(v); });
        // At this point, the capsule object is created successfully.
        // Release the unique_ptr and let the capsule object own the storage to avoid double-free.
        (void) storage_ptr.release();

        // Use `PyDict_SetDefault` for atomic test-and-set:
        //   - If key doesn't exist, inserts our capsule and returns it.
        //   - If key exists (another thread inserted first), returns the existing value.
        // This is thread-safe because `PyDict_SetDefault` will hold a lock on the dict.
        //
        // NOTE: Here we use `PyDict_SetDefault` instead of `PyDict_SetDefaultRef` because the
        //       capsule is kept alive until interpreter shutdown, so we do not need to handle
        //       incref and decref here.
        capsule_obj = dict_setdefaultstring(state_dict.ptr(), key, new_capsule.ptr());
        if (capsule_obj == nullptr) {
            throw error_already_set();
        }
        created = (capsule_obj == new_capsule.ptr());
        // - If key already existed, our `new_capsule` is not inserted, it will be destructed when
        //   going out of scope here, and will call the destructor set above.
        // - Otherwise, our `new_capsule` is now in the dict, and it owns the storage and the state
        //   dict will incref it.  We need to set the caller's destructor on it, which will be
        //   called when the interpreter shuts down.
        if (created && dtor != PYBIND11_DTOR_USE_DELETE) {
            if (PyCapsule_SetDestructor(capsule_obj, dtor) < 0) {
                throw error_already_set();
            }
        }
    }

    // Get the storage pointer from the capsule.
    void *raw_ptr = PyCapsule_GetPointer(capsule_obj, /*name=*/nullptr);
    if (!raw_ptr) {
        raise_from(PyExc_SystemError,
                   "pybind11::detail::atomic_get_or_create_in_state_dict() FAILED");
        throw error_already_set();
    }
    return std::pair<Payload *, bool>(static_cast<Payload *>(raw_ptr), created);
}

#undef PYBIND11_DTOR_USE_DELETE

template <typename InternalsType>
class internals_pp_manager {
public:
    using on_fetch_function = void(InternalsType *);

    static internals_pp_manager &get_instance(char const *id, on_fetch_function *on_fetch) {
        static internals_pp_manager instance(id, on_fetch);
        return instance;
    }

    /// Get the current pointer-to-pointer, allocating it if it does not already exist.  May
    /// acquire the GIL. Will never return nullptr.
    std::unique_ptr<InternalsType> *get_pp() {
#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
        if (has_seen_non_main_interpreter()) {
            // 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.
            auto *tstate = get_thread_state_unchecked();
            if (!tstate || tstate->interp != last_istate_tls()) {
                gil_scoped_acquire_simple gil;
                if (!tstate) {
                    tstate = get_thread_state_unchecked();
                }
                last_istate_tls() = tstate->interp;
                internals_p_tls() = get_or_create_pp_in_state_dict();
            }
            return internals_p_tls();
        }
#endif
        if (!internals_singleton_pp_) {
            gil_scoped_acquire_simple gil;
            internals_singleton_pp_ = get_or_create_pp_in_state_dict();
        }
        return internals_singleton_pp_;
    }

    /// Drop all the references we're currently holding.
    void unref() {
#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
        if (has_seen_non_main_interpreter()) {
            last_istate_tls() = nullptr;
            internals_p_tls() = nullptr;
            return;
        }
#endif
        internals_singleton_pp_ = nullptr;
    }

    void destroy() {
#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
        if (has_seen_non_main_interpreter()) {
            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()) {
                auto tpp = internals_p_tls();
                {
                    std::lock_guard<std::mutex> lock(pp_set_mutex_);
                    pps_have_created_content_.erase(tpp); // untrack deleted pp
                }
                delete tpp; // may call back into Python
            }
            unref();
            return;
        }
#endif
        {
            std::lock_guard<std::mutex> lock(pp_set_mutex_);
            pps_have_created_content_.erase(internals_singleton_pp_); // untrack deleted pp
        }
        delete internals_singleton_pp_; // may call back into Python
        unref();
    }

    void create_pp_content_once(std::unique_ptr<InternalsType> *const pp) {
        // Assume the GIL is held here. May call back into Python. We cannot hold the lock with our
        // mutex here. So there may be multiple threads creating the content at the same time. Only
        // one will install its content to pp below. Others will be freed when going out of scope.
        auto tmp = std::unique_ptr<InternalsType>(new InternalsType());

        {
            // Lock scope must not include Python calls, which may require the GIL and cause
            // deadlocks.
            std::lock_guard<std::mutex> lock(pp_set_mutex_);

            if (*pp) {
                // Already created in another thread.
                return;
            }

            // At this point, pp->get() is nullptr.
            // The content is either not yet created, or was previously destroyed via pp->reset().

            // Detect re-creation of internals after destruction during interpreter shutdown.
            // If pybind11 code (e.g., tp_traverse/tp_clear calling py::cast) runs after internals
            // have been destroyed, a new empty internals would be created, causing type lookup
            // failures. See also get_or_create_pp_in_state_dict() comments.
            if (pps_have_created_content_.find(pp) != pps_have_created_content_.end()) {
                pybind11_fail(
                    "pybind11::detail::internals_pp_manager::create_pp_content_once() "
                    "FAILED: reentrant call detected while fetching pybind11 internals!");
            }

            // Each interpreter can only create its internals once.
            pps_have_created_content_.insert(pp);
            // Install the created content.
            pp->swap(tmp);
        }
    }

private:
    internals_pp_manager(char const *id, on_fetch_function *on_fetch)
        : holder_id_(id), on_fetch_(on_fetch) {}

    std::unique_ptr<InternalsType> *get_or_create_pp_in_state_dict() {
        // The `unique_ptr<InternalsType>` is intentionally leaked on interpreter shutdown.
        // Once an instance is created, it will never be deleted until the process exits (compare
        // to interpreter shutdown in multiple-interpreter scenarios).
        // We cannot guarantee the destruction order of capsules in the interpreter state dict on
        // interpreter shutdown, so deleting internals too early could cause undefined behavior
        // when other pybind11 objects access `get_internals()` during finalization (which would
        // recreate empty internals). See also create_pp_content_once() above.
        // See https://github.com/pybind/pybind11/pull/5958#discussion_r2717645230.
        auto result = atomic_get_or_create_in_state_dict<std::unique_ptr<InternalsType>>(
            holder_id_, /*dtor=*/nullptr /* leak the capsule content */);
        auto *pp = result.first;
        bool created = result.second;
        // Only call on_fetch_ when fetching existing internals, not when creating new ones.
        if (!created && on_fetch_ && pp) {
            on_fetch_(pp->get());
        }
        return pp;
    }

#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
    static PyInterpreterState *&last_istate_tls() {
        static thread_local PyInterpreterState *last_istate = nullptr;
        return last_istate;
    }

    static std::unique_ptr<InternalsType> *&internals_p_tls() {
        static thread_local std::unique_ptr<InternalsType> *internals_p = nullptr;
        return internals_p;
    }
#endif

    char const *holder_id_ = nullptr;
    on_fetch_function *on_fetch_ = nullptr;
    // Pointer-to-pointer to the singleton internals for the first seen interpreter (may not be the
    // main interpreter)
    std::unique_ptr<InternalsType> *internals_singleton_pp_ = nullptr;

    // Track pointer-to-pointers whose internals have been created, to detect re-entrancy.
    // Use instance member over static due to singleton pattern of this class.
    std::unordered_set<std::unique_ptr<InternalsType> *> pps_have_created_content_;
    std::mutex pp_set_mutex_;
};

// If We loaded the internals through `state_dict`, our `error_already_set`
// and `builtin_exception` may be different local classes than the ones set up in the
// initial exception translator, below, so add another for our local exception classes.
//
// libstdc++ doesn't require this (types there are identified only by name)
// libc++ with CPython doesn't require this (types are explicitly exported)
// libc++ with PyPy still need it, awaiting further investigation
#if !defined(__GLIBCXX__)
inline void check_internals_local_exception_translator(internals *internals_ptr) {
    if (internals_ptr) {
        for (auto et : internals_ptr->registered_exception_translators) {
            if (et == &translate_local_exception) {
                return;
            }
        }
        internals_ptr->registered_exception_translators.push_front(&translate_local_exception);
    }
}
#endif

inline internals_pp_manager<internals> &get_internals_pp_manager() {
#if defined(__GLIBCXX__)
#    define ON_FETCH_FN nullptr
#else
#    define ON_FETCH_FN &check_internals_local_exception_translator
#endif
    return internals_pp_manager<internals>::get_instance(PYBIND11_INTERNALS_ID, ON_FETCH_FN);
#undef ON_FETCH_FN
}

/// Return a reference to the current `internals` data
PYBIND11_NOINLINE internals &get_internals() {
    auto &ppmgr = get_internals_pp_manager();
    auto *pp = ppmgr.get_pp();
    if (!pp) {
        pybind11_fail("get_internals: get_pp() returned nullptr");
    }
    auto &internals_ptr = *pp;
    if (!internals_ptr) {
        // Slow path, something needs fetched from the state dict or created
        gil_scoped_acquire_simple gil;
        error_scope err_scope;

        ppmgr.create_pp_content_once(&internals_ptr);

        if (!internals_ptr) {
            pybind11_fail("get_internals: create_pp_content_once() produced nullptr");
        }
        if (!internals_ptr->instance_base) {
            // This calls get_internals, so cannot be called from within the internals constructor
            // called above because internals_ptr must be set before get_internals is called again
            internals_ptr->instance_base = make_object_base_type(internals_ptr->default_metaclass);
        }
    }
    return *internals_ptr;
}

/// Return the PyObject* for the internals capsule (borrowed reference).
/// Returns nullptr if the capsule doesn't exist yet.
inline PyObject *get_internals_capsule() {
    auto state_dict = reinterpret_borrow<dict>(get_python_state_dict());
    return dict_getitemstring(state_dict.ptr(), PYBIND11_INTERNALS_ID);
}

/// Return the key used for local_internals in the state dict.
/// This function ensures a consistent key is used across all call sites within the same
/// compilation unit. The key includes the address of a static variable to make it unique per
/// module (DSO), matching the behavior of get_local_internals_pp_manager().
inline const std::string &get_local_internals_key() {
    static const std::string key
        = PYBIND11_MODULE_LOCAL_ID + std::to_string(reinterpret_cast<uintptr_t>(&key));
    return key;
}

/// Return the PyObject* for the local_internals capsule (borrowed reference).
/// Returns nullptr if the capsule doesn't exist yet.
inline PyObject *get_local_internals_capsule() {
    const auto &key = get_local_internals_key();
    auto state_dict = reinterpret_borrow<dict>(get_python_state_dict());
    return dict_getitemstring(state_dict.ptr(), key.c_str());
}

inline void ensure_internals() {
    pybind11::detail::get_internals_pp_manager().unref();
#ifdef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
    if (PyInterpreterState_Get() != PyInterpreterState_Main()) {
        has_seen_non_main_interpreter() = true;
    }
#endif
    pybind11::detail::get_internals();
}

inline internals_pp_manager<local_internals> &get_local_internals_pp_manager() {
    // Use the address of a static variable as part of the key, so that the value is uniquely tied
    // to where the module is loaded in memory
    return internals_pp_manager<local_internals>::get_instance(get_local_internals_key().c_str(),
                                                               nullptr);
}

/// Works like `get_internals`, but for things which are locally registered.
inline local_internals &get_local_internals() {
    auto &ppmgr = get_local_internals_pp_manager();
    auto &internals_ptr = *ppmgr.get_pp();
    if (!internals_ptr) {
        gil_scoped_acquire_simple gil;
        error_scope err_scope;

        ppmgr.create_pp_content_once(&internals_ptr);
    }
    return *internals_ptr;
}

#ifdef Py_GIL_DISABLED
#    define PYBIND11_LOCK_INTERNALS(internals) pycritical_section 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);
}

template <typename F>
inline void with_internals_if_internals(const F &cb) {
    auto &ppmgr = get_internals_pp_manager();
    auto &internals_ptr = *ppmgr.get_pp();
    if (internals_ptr) {
        auto &internals = *internals_ptr;
        PYBIND11_LOCK_INTERNALS(internals);
        cb(internals);
    }
}

template <typename F>
inline auto with_exception_translators(const F &cb)
    -> decltype(cb(get_internals().registered_exception_translators,
                   get_local_internals().registered_exception_translators)) {
    auto &internals = get_internals();
#ifdef Py_GIL_DISABLED
    pycritical_section lock((internals).exception_translator_mutex);
#endif
    auto &local_internals = get_local_internals();
    return cb(internals.registered_exception_translators,
              local_internals.registered_exception_translators);
}

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<pymutex> 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<pymutex> 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) {
    // 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();
}

PYBIND11_NAMESPACE_END(detail)

/// Returns a named pointer that is shared among all extension modules (using the same
/// 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) {
    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) {
    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
/// such entry exists. Otherwise, a new object of default-constructible type `T` is
/// 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) {
    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;
    });
}

PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)