Add per-interpreter storage for gil_safe_call_once_and_store

include/pybind11/detail/internals.h

@@ -660,6 +660,8 @@ class internals_pp_manager {
 
     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_;
 };
 

include/pybind11/gil_safe_call_once.h

@@ -3,13 +3,18 @@
 #pragma once
 
 #include "detail/common.h"
+#include "detail/internals.h"
 #include "gil.h"
 
 #include <cassert>
 #include <mutex>
 
-#ifdef Py_GIL_DISABLED
+#if defined(Py_GIL_DISABLED) || defined(PYBIND11_HAS_SUBINTERPRETER_SUPPORT)
 #    include <atomic>
+
+using atomic_bool = std::atomic_bool;
+#else
+using atomic_bool = bool;
 #endif
 
 PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
@@ -48,12 +53,19 @@ PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
 // functions, which is usually the case.
 //
 // For in-depth background, see docs/advanced/deadlock.md
+#ifndef PYBIND11_HAS_SUBINTERPRETER_SUPPORT
+// Subinterpreter support is disabled.
+// In this case, we can store the result globally, because there is only a single interpreter.
+//
+// The life span of the stored result is the entire process lifetime. It is leaked on process
+// termination to avoid destructor calls after the Python interpreter was finalized.
 template <typename T>
 class gil_safe_call_once_and_store {
 public:
     // PRECONDITION: The GIL must be held when `call_once_and_store_result()` is called.
     template <typename Callable>
-    gil_safe_call_once_and_store &call_once_and_store_result(Callable &&fn) {
+    gil_safe_call_once_and_store &call_once_and_store_result(Callable &&fn,
+                                                             void (*)(T &) /*unused*/ = nullptr) {
         if (!is_initialized_) { // This read is guarded by the GIL.
             // Multiple threads may enter here, because the GIL is released in the next line and
             // CPython API calls in the `fn()` call below may release and reacquire the GIL.
@@ -74,29 +86,187 @@ class gil_safe_call_once_and_store {
     T &get_stored() {
         assert(is_initialized_);
         PYBIND11_WARNING_PUSH
-#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5
+#    if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5
         // Needed for gcc 4.8.5
         PYBIND11_WARNING_DISABLE_GCC("-Wstrict-aliasing")
-#endif
+#    endif
         return *reinterpret_cast<T *>(storage_);
         PYBIND11_WARNING_POP
     }
 
     constexpr gil_safe_call_once_and_store() = default;
+    // The instance is a global static, so its destructor runs when the process
+    // is terminating. Therefore, do nothing here because the Python interpreter
+    // may have been finalized already.
     PYBIND11_DTOR_CONSTEXPR ~gil_safe_call_once_and_store() = default;
 
 private:
+    // Global static storage (per process) when subinterpreter support is disabled.
     alignas(T) char storage_[sizeof(T)] = {};
     std::once_flag once_flag_;
-#ifdef Py_GIL_DISABLED
-    std::atomic_bool
-#else
-    bool
-#endif
-        is_initialized_{false};
+
     // The `is_initialized_`-`storage_` pair is very similar to `std::optional`,
     // but the latter does not have the triviality properties of former,
     // therefore `std::optional` is not a viable alternative here.
+    atomic_bool is_initialized_{false};
+};
+#else
+// Subinterpreter support is enabled.
+// In this case, we should store the result per-interpreter instead of globally, because each
+// subinterpreter has its own separate state. The cached result may not shareable across
+// interpreters (e.g., imported modules and their members).
+
+struct call_once_storage_base {
+    call_once_storage_base() = default;
+    virtual ~call_once_storage_base() = default;
+    call_once_storage_base(const call_once_storage_base &) = delete;
+    call_once_storage_base(call_once_storage_base &&) = delete;
+    call_once_storage_base &operator=(const call_once_storage_base &) = delete;
+    call_once_storage_base &operator=(call_once_storage_base &&) = delete;
+};
+
+template <typename T>
+struct call_once_storage : call_once_storage_base {
+    alignas(T) char storage[sizeof(T)] = {};
+    std::once_flag once_flag;
+    void (*finalize)(T &) = nullptr;
+    std::atomic_bool is_initialized{false};
+
+    call_once_storage() = default;
+    ~call_once_storage() override {
+        if (is_initialized) {
+            if (finalize != nullptr) {
+                finalize(*reinterpret_cast<T *>(storage));
+            } else {
+                reinterpret_cast<T *>(storage)->~T();
+            }
+        }
+    };
+    call_once_storage(const call_once_storage &) = delete;
+    call_once_storage(call_once_storage &&) = delete;
+    call_once_storage &operator=(const call_once_storage &) = delete;
+    call_once_storage &operator=(call_once_storage &&) = delete;
 };
 
+/// Storage map for `gil_safe_call_once_and_store`. Stored in a capsule in the interpreter's state
+/// dict with proper destructor to ensure cleanup when the interpreter is destroyed.
+using call_once_storage_map_type = std::unordered_map<const void *, call_once_storage_base *>;
+
+#    define PYBIND11_CALL_ONCE_STORAGE_MAP_ID PYBIND11_INTERNALS_ID "_call_once_storage_map__"
+
+// The life span of the stored result is the entire interpreter lifetime. An additional
+// `finalize_fn` can be provided to clean up the stored result when the interpreter is destroyed.
+template <typename T>
+class gil_safe_call_once_and_store {
+public:
+    // PRECONDITION: The GIL must be held when `call_once_and_store_result()` is called.
+    template <typename Callable>
+    gil_safe_call_once_and_store &call_once_and_store_result(Callable &&fn,
+                                                             void (*finalize_fn)(T &) = nullptr) {
+        if (!is_last_storage_valid()) {
+            // Multiple threads may enter here, because the GIL is released in the next line and
+            // CPython API calls in the `fn()` call below may release and reacquire the GIL.
+            gil_scoped_release gil_rel; // Needed to establish lock ordering.
+            const void *const key = reinterpret_cast<const void *>(this);
+            // There can be multiple threads going through here.
+            call_once_storage<T> *value = nullptr;
+            {
+                gil_scoped_acquire gil_acq;
+                // Only one thread will enter here at a time.
+                auto &storage_map = *get_or_create_call_once_storage_map();
+                const auto it = storage_map.find(key);
+                if (it != storage_map.end()) {
+                    value = static_cast<call_once_storage<T> *>(it->second);
+                } else {
+                    value = new call_once_storage<T>{};
+                    storage_map.emplace(key, value);
+                }
+            }
+            assert(value != nullptr);
+            std::call_once(value->once_flag, [&] {
+                // Only one thread will ever enter here.
+                gil_scoped_acquire gil_acq;
+                // fn may release, but will reacquire, the GIL.
+                ::new (value->storage) T(fn());
+                value->finalize = finalize_fn;
+                value->is_initialized = true;
+                last_storage_ptr_ = reinterpret_cast<T *>(value->storage);
+                is_initialized_by_atleast_one_interpreter_ = true;
+            });
+            // All threads will observe `is_initialized_by_atleast_one_interpreter_` as true here.
+        }
+        // Intentionally not returning `T &` to ensure the calling code is self-documenting.
+        return *this;
+    }
+
+    // This must only be called after `call_once_and_store_result()` was called.
+    T &get_stored() {
+        T *result = last_storage_ptr_;
+        if (!is_last_storage_valid()) {
+            gil_scoped_acquire gil_acq;
+            const void *const key = reinterpret_cast<const void *>(this);
+            auto &storage_map = *get_or_create_call_once_storage_map();
+            auto *value = static_cast<call_once_storage<T> *>(storage_map.at(key));
+            result = last_storage_ptr_ = reinterpret_cast<T *>(value->storage);
+        }
+        assert(result != nullptr);
+        return *result;
+    }
+
+    gil_safe_call_once_and_store() = default;
+    // The instance is a global static, so its destructor runs when the process
+    // is terminating. Therefore, do nothing here because the Python interpreter
+    // may have been finalized already.
+    PYBIND11_DTOR_CONSTEXPR ~gil_safe_call_once_and_store() = default;
+
+private:
+    bool is_last_storage_valid() const {
+        return is_initialized_by_atleast_one_interpreter_
+               && detail::get_num_interpreters_seen() == 1;
+    }
+
+    static call_once_storage_map_type *get_or_create_call_once_storage_map() {
+        error_scope err_scope;
+        dict state_dict = detail::get_python_state_dict();
+        auto storage_map_obj = reinterpret_steal<object>(
+            detail::dict_getitemstringref(state_dict.ptr(), PYBIND11_CALL_ONCE_STORAGE_MAP_ID));
+        call_once_storage_map_type *storage_map = nullptr;
+        if (storage_map_obj) {
+            void *raw_ptr = PyCapsule_GetPointer(storage_map_obj.ptr(), /*name=*/nullptr);
+            if (!raw_ptr) {
+                raise_from(PyExc_SystemError,
+                           "pybind11::gil_safe_call_once_and_store::"
+                           "get_or_create_call_once_storage_map() FAILED");
+                throw error_already_set();
+            }
+            storage_map = reinterpret_cast<call_once_storage_map_type *>(raw_ptr);
+        } else {
+            storage_map = new call_once_storage_map_type();
+            // Create capsule with destructor to clean up the storage map when the interpreter
+            // shuts down
+            state_dict[PYBIND11_CALL_ONCE_STORAGE_MAP_ID]
+                = capsule(storage_map, [](void *ptr) noexcept {
+                      auto *map = reinterpret_cast<call_once_storage_map_type *>(ptr);
+                      for (const auto &entry : *map) {
+                          delete entry.second;
+                      }
+                      delete map;
+                  });
+        }
+        return storage_map;
+    }
+
+    // No storage needed when subinterpreter support is enabled.
+    // The actual storage is stored in the per-interpreter state dict via
+    // `get_or_create_call_once_storage_map()`.
+
+    // Fast local cache to avoid repeated lookups when there are no multiple interpreters.
+    // This is only valid if there is a single interpreter. Otherwise, it is not used.
+    T *last_storage_ptr_ = nullptr;
+    // This flag is true if the value has been initialized by any interpreter (may not be the
+    // current one).
+    atomic_bool is_initialized_by_atleast_one_interpreter_{false};
+};
+#endif
+
 PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)