FazBrowse GitHub Viewer | Trending |
URL:
| Home
Tools: [Download Repo ZIP]   [Original HTTPS Page]

GitHub Viewer

#include #include #include #include #include #include #include #include #include #include #include #include "python.hpp" extern "C" void pythonx_handle_io_write(const char *message, const char *eval_info_bytes, bool type); namespace pythonx { using namespace python; // State std::mutex init_mutex; bool is_initialized = false; std::wstring python_home_path_w; std::map compilation_cache; std::mutex compilation_cache_mutex; // Wrapper around the Python Global Interpreter Lock (GIL). // // To acquire the GIL, the caller simply needs to initialize a new // guard object. Once the guard object's lifetime ends, the GIL is // automatically released. This is the use of RAII [1] pattern, // similarly to `std::lock_guard`. // // [1]: https://en.wikipedia.org/wiki/Resource_acquisition_is_initialization class PyGILGuard { PyGILState_STATE state; public: PyGILGuard() { this->state = PyGILState_Ensure(); } ~PyGILGuard() { PyGILState_Release(this->state); } }; // Ensures the given object refcount is decremented when the guard // goes out of scope. class PyDecRefGuard { PyObjectPtr py_object; public: PyDecRefGuard() : py_object(nullptr) {} PyDecRefGuard(PyObjectPtr py_object) : py_object(py_object) {} ~PyDecRefGuard() { if (this->py_object != nullptr) { Py_DecRef(this->py_object); } } PyDecRefGuard &operator=(PyObjectPtr py_object) { this->py_object = py_object; return *this; } }; void ensure_initialized() { auto init_guard = std::lock_guard(init_mutex); if (!is_initialized) { throw std::runtime_error("Python interpreter has not been initialized"); } } namespace atoms { auto ElixirPythonxError = fine::Atom("Elixir.Pythonx.Error"); auto ElixirPythonxJanitor = fine::Atom("Elixir.Pythonx.Janitor"); auto ElixirPythonxObject = fine::Atom("Elixir.Pythonx.Object"); auto decref = fine::Atom("decref"); auto integer = fine::Atom("integer"); auto list = fine::Atom("list"); auto map = fine::Atom("map"); auto map_set = fine::Atom("map_set"); auto output = fine::Atom("output"); auto ref = fine::Atom("ref"); auto traceback = fine::Atom("traceback"); auto tuple = fine::Atom("tuple"); auto type = fine::Atom("type"); auto value = fine::Atom("value"); } // namespace atoms struct ExObjectResource { PyObjectPtr py_object; ExObjectResource(PyObjectPtr py_object) : py_object(py_object) {} void destructor(ErlNifEnv *env) { // Decrementing refcount requires GIL and we should not block in // the destructor, so we send a message to a known process and let // it decrement the refcount for us. Also see [1]. // // [1]:https://erlangforums.com/t/how-to-deal-with-destructors-that-can-take-a-while-to-run-and-possibly-block-the-scheduler/4290 if (!is_initialized) { // If we allow multiple initializations, we need to add a counter // and check that py_object comes from the current initialization return; } auto ptr = reinterpret_cast(this->py_object); auto janitor_name = fine::encode(env, atoms::ElixirPythonxJanitor); ErlNifPid janitor_pid; if (enif_whereis_pid(env, janitor_name, &janitor_pid)) { auto msg_env = enif_alloc_env(); auto msg = fine::encode(msg_env, std::make_tuple(atoms::decref, ptr)); enif_send(env, &janitor_pid, msg_env, msg); enif_free_env(msg_env); } else { std::cerr py_object, ex_key.resource->py_object, ex_value.resource->py_object); raise_if_failed(env, result); return fine::Ok(); } FINE_NIF(dict_set_item, ERL_NIF_DIRTY_JOB_CPU_BOUND); ExObject tuple_new(ErlNifEnv *env, uint64_t size) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_tuple = PyTuple_New(size); raise_if_failed(env, py_tuple); return ExObject(fine::make_resource(py_tuple)); } FINE_NIF(tuple_new, ERL_NIF_DIRTY_JOB_CPU_BOUND); fine::Ok tuple_set_item(ErlNifEnv *env, ExObject ex_object, uint64_t index, ExObject ex_value) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto result = PyTuple_SetItem(ex_object.resource->py_object, index, ex_value.resource->py_object); raise_if_failed(env, result); // PyTuple_SetItem steals a reference, so we add one back Py_IncRef(ex_value.resource->py_object); return fine::Ok(); } FINE_NIF(tuple_set_item, ERL_NIF_DIRTY_JOB_CPU_BOUND); ExObject list_new(ErlNifEnv *env, uint64_t size) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_tuple = PyList_New(size); raise_if_failed(env, py_tuple); return ExObject(fine::make_resource(py_tuple)); } FINE_NIF(list_new, ERL_NIF_DIRTY_JOB_CPU_BOUND); fine::Ok list_set_item(ErlNifEnv *env, ExObject ex_object, uint64_t index, ExObject ex_value) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto result = PyList_SetItem(ex_object.resource->py_object, index, ex_value.resource->py_object); raise_if_failed(env, result); // PyList_SetItem steals a reference, so we add one back Py_IncRef(ex_value.resource->py_object); return fine::Ok(); } FINE_NIF(list_set_item, ERL_NIF_DIRTY_JOB_CPU_BOUND); ExObject set_new(ErlNifEnv *env) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_set = PySet_New(NULL); raise_if_failed(env, py_set); return ExObject(fine::make_resource(py_set)); } FINE_NIF(set_new, ERL_NIF_DIRTY_JOB_CPU_BOUND); fine::Ok set_add(ErlNifEnv *env, ExObject ex_object, ExObject ex_key) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto result = PySet_Add(ex_object.resource->py_object, ex_key.resource->py_object); raise_if_failed(env, result); return fine::Ok(); } FINE_NIF(set_add, ERL_NIF_DIRTY_JOB_CPU_BOUND); ExObject object_repr(ErlNifEnv *env, ExObject ex_object) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_repr = PyObject_Repr(ex_object.resource->py_object); raise_if_failed(env, py_repr); return ExObject(fine::make_resource(py_repr)); } FINE_NIF(object_repr, ERL_NIF_DIRTY_JOB_CPU_BOUND); fine::Term format_exception(ErlNifEnv *env, ExError error) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_traceback_module = PyImport_ImportModule("traceback"); raise_if_failed(env, py_traceback_module); auto py_traceback_module_guard = PyDecRefGuard(py_traceback_module); auto format_exception = PyObject_GetAttrString(py_traceback_module, "format_exception"); raise_if_failed(env, format_exception); auto format_exception_guard = PyDecRefGuard(format_exception); auto py_type = error.type.resource->py_object; auto py_value = error.value.resource->py_object; auto py_traceback = error.traceback.resource->py_object; auto format_exception_args = PyTuple_Pack(3, py_type, py_value, py_traceback); raise_if_failed(env, format_exception_args); auto format_exception_args_guard = PyDecRefGuard(format_exception_args); auto py_lines = PyObject_Call(format_exception, format_exception_args, NULL); raise_if_failed(env, py_lines); auto py_lines_guard = PyDecRefGuard(py_lines); auto size = PyList_Size(py_lines); raise_if_failed(env, size); auto terms = std::vector(); terms.reserve(size); for (Py_ssize_t i = 0; i < size; i++) { auto py_line = PyList_GetItem(py_lines, i); raise_if_failed(env, py_line); terms.push_back(py_object_to_binary_term(env, py_line)); } return enif_make_list_from_array(env, terms.data(), static_cast(size)); } FINE_NIF(format_exception, ERL_NIF_DIRTY_JOB_CPU_BOUND); fine::Term decode_once(ErlNifEnv *env, ExObject ex_object) { ensure_initialized(); auto gil_guard = PyGILGuard(); auto py_object = ex_object.resource->py_object; auto is_none = Py_IsNone(py_object); raise_if_failed(env, is_none); if (is_none) { return fine::encode(env, std::nullopt); } auto is_true = Py_IsTrue(py_object); raise_if_failed(env, is_true); if (is_true) { return fine::encode(env, true); } auto is_false = Py_IsFalse(py_object); raise_if_failed(env, is_false); if (is_false) { return fine::encode(env, false); } auto py_builtins = PyEval_GetBuiltins(); raise_if_failed(env, py_builtins); auto py_int_type = PyDict_GetItemString(py_builtins, "int"); raise_if_failed(env, py_int_type); auto is_long = PyObject_IsInstance(py_object, py_int_type); raise_if_failed(env, is_long); if (is_long) { int overflow; auto integer = PyLong_AsLongLongAndOverflow(py_object, &overflow); if (PyErr_Occurred() != NULL) { raise_py_error(env); } if (overflow == 0) { return enif_make_int64(env, integer); } // Integer over 64 bits auto py_str = PyObject_Str(py_object); raise_if_failed(env, py_str); auto py_str_guard = PyDecRefGuard(py_str); auto binary_term = py_object_to_binary_term(env, py_str); return fine::encode( env, std::make_tuple(atoms::integer, fine::Term(binary_term))); } auto py_float_type = PyDict_GetItemString(py_builtins, "float"); raise_if_failed(env, py_float_type); auto is_float = PyObject_IsInstance(py_object, py_float_type); raise_if_failed(env, is_float); if (is_float) { double number = PyFloat_AsDouble(py_object); if (PyErr_Occurred() != NULL) { raise_py_error(env); } return enif_make_double(env, number); } auto py_tuple_type = PyDict_GetItemString(py_builtins, "tuple"); raise_if_failed(env, py_tuple_type); auto is_tuple = PyObject_IsInstance(py_object, py_tuple_type); raise_if_failed(env, is_tuple); if (is_tuple) { auto size = PyTuple_Size(py_object); raise_if_failed(env, size); auto terms = std::vector(); terms.reserve(size); for (Py_ssize_t i = 0; i < size; i++) { auto py_item = PyTuple_GetItem(py_object, i); raise_if_failed(env, py_item); Py_IncRef(py_item); auto ex_item = ExObject(fine::make_resource(py_item)); terms.push_back(fine::encode(env, ex_item)); } auto items = enif_make_list_from_array(env, terms.data(), static_cast(size)); return fine::encode(env, std::make_tuple(atoms::tuple, fine::Term(items))); } auto py_list_type = PyDict_GetItemString(py_builtins, "list"); raise_if_failed(env, py_list_type); auto is_list = PyObject_IsInstance(py_object, py_list_type); raise_if_failed(env, is_list); if (is_list) { auto size = PyList_Size(py_object); raise_if_failed(env, size); auto terms = std::vector(); terms.reserve(size); for (Py_ssize_t i = 0; i < size; i++) { auto py_item = PyList_GetItem(py_object, i); raise_if_failed(env, py_item); Py_IncRef(py_item); auto ex_item = ExObject(fine::make_resource(py_item)); terms.push_back(fine::encode(env, ex_item)); } auto items = enif_make_list_from_array(env, terms.data(), static_cast(size)); return fine::encode(env, std::make_tuple(atoms::list, fine::Term(items))); } auto py_dict_type = PyDict_GetItemString(py_builtins, "dict"); raise_if_failed(env, py_dict_type); auto is_dict = PyObject_IsInstance(py_object, py_dict_type); raise_if_failed(env, is_dict); if (is_dict) { auto size = PyDict_Size(py_object); raise_if_failed(env, size); auto terms = std::vector(); terms.reserve(size); PyObjectPtr py_key, py_value; Py_ssize_t pos = 0; while (PyDict_Next(py_object, &pos, &py_key, &py_value)) { Py_IncRef(py_key); auto ex_key = ExObject(fine::make_resource(py_key)); Py_IncRef(py_value); auto ex_value = ExObject(fine::make_resource(py_value)); terms.push_back(fine::encode(env, std::make_tuple(ex_key, ex_value))); } auto items = enif_make_list_from_array(env, terms.data(), static_cast(size)); return fine::encode(env, std::make_tuple(atoms::map, fine::Term(items))); } auto py_str_type = PyDict_GetItemString(py_builtins, "str"); raise_if_failed(env, py_str_type); auto is_unicode = PyObject_IsInstance(py_object, py_str_type); raise_if_failed(env, is_unicode); if (is_unicode) { return py_object_to_binary_term(env, py_object); } auto py_bytes_type = PyDict_GetItemString(py_builtins, "bytes"); raise_if_failed(env, py_bytes_type); auto is_bytes = PyObject_IsInstance(py_object, py_bytes_type); raise_if_failed(env, is_bytes); if (is_bytes) { Py_ssize_t size; char *buffer; auto result = PyBytes_AsStringAndSize(py_object, &buffer, &size); raise_if_failed(env, result); ERL_NIF_TERM binary_term; auto data = enif_make_new_binary(env, size, &binary_term); memcpy(data, buffer, size); return binary_term; } auto py_set_type = PyDict_GetItemString(py_builtins, "set"); raise_if_failed(env, py_set_type); auto is_set = PyObject_IsInstance(py_object, py_set_type); raise_if_failed(env, is_set); auto py_frozenset_type = PyDict_GetItemString(py_builtins, "frozenset"); raise_if_failed(env, py_frozenset_type); auto is_frozenset = PyObject_IsInstance(py_object, py_frozenset_type); raise_if_failed(env, is_frozenset); if (is_set || is_frozenset) { auto size = PySet_Size(py_object); raise_if_failed(env, size); auto terms = std::vector(); terms.reserve(size); auto py_iter = PyObject_GetIter(py_object); raise_if_failed(env, py_iter); auto py_iter_guard = PyDecRefGuard(py_iter); PyObjectPtr py_item = NULL; while ((py_item = PyIter_Next(py_iter)) != NULL) { // Note that PyIter_Next already returns a new reference auto ex_item = ExObject(fine::make_resource(py_item)); terms.push_back(fine::encode(env, ex_item)); } if (PyErr_Occurred() != NULL) { raise_py_error(env); } auto items = enif_make_list_from_array(env, terms.data(), static_cast(size)); return fine::encode(env, std::make_tuple(atoms::map_set, fine::Term(items))); } // None of the built-ins, return %Pythonx.Object{} as is return fine::encode(env, ex_object); } FINE_NIF(decode_once, ERL_NIF_DIRTY_JOB_CPU_BOUND); std::tuple compile(ErlNifEnv *env, ErlNifBinary code) { // Python code can be compiled in either "exec" mode (multiple // statements with no result value) or "eval" mode (single expression // with a result value). We want our eval API to accept arbitrary // Python code with multiple statements, while also returning the // final result. To achieve this we parse the code using Python // standard library and check if the last statement is an expression. // If that is the case, we split the code, "exec" the statements // and "eval" the final expression separately. // // For the reference, below is a Python code corresponding to the // described logic. Technically we could "exec" that Python code, // but in order to avoid extra overhead we call the Python functions // directly via the C API. // // import ast // // module = ast.parse(code, "", mode="exec") // // body_code = None // last_expr_code = None // // if module.body: // last_statement = module.body[-1] // // if isinstance(last_statement, ast.Expr): // expr = ast.Expression(module.body.pop().value) // # Copy positional information to the expression root node // expr.lineno = last_statement.lineno // expr.col_offset = last_statement.col_offset // expr.end_col_offset = last_statement.end_col_offset // last_expr_code = compile(expr, filename, mode="eval") // // if module.body: // body_code = compile(module, filename, mode="exec") // // The body and last expression is then evaluated separately, as in: // // if body_code: // eval(body_code) // // if last_expr_code: // result = eval(last_expr_code) // else: // result = None PyObjectPtr py_body_code = nullptr; PyObjectPtr py_last_expr_code = nullptr; auto py_last_expr_code_guard = PyDecRefGuard(); auto py_ast = PyImport_ImportModule("ast"); raise_if_failed(env, py_ast); auto py_ast_guard = PyDecRefGuard(py_ast); auto py_parse = PyObject_GetAttrString(py_ast, "parse"); raise_if_failed(env, py_parse); auto py_parse_guard = PyDecRefGuard(py_parse); auto py_code = PyUnicode_FromStringAndSize( reinterpret_cast(code.data), code.size); raise_if_failed(env, py_code); auto py_code_guard = PyDecRefGuard(py_code); auto py_file_string = PyUnicode_FromStringAndSize("", 8); raise_if_failed(env, py_file_string); auto py_file_string_guard = PyDecRefGuard(py_file_string); auto py_exec_string = PyUnicode_FromStringAndSize("exec", 4); raise_if_failed(env, py_exec_string); auto py_exec_string_guard = PyDecRefGuard(py_exec_string); auto py_parse_args = PyTuple_Pack(3, py_code, py_file_string, py_exec_string); raise_if_failed(env, py_parse_args); auto py_parse_args_guard = PyDecRefGuard(py_parse_args); auto py_module_ast = PyObject_Call(py_parse, py_parse_args, NULL); raise_if_failed(env, py_module_ast); auto py_module_ast_guard = PyDecRefGuard(py_module_ast); auto py_builtins = PyEval_GetBuiltins(); raise_if_failed(env, py_builtins); auto py_compile = PyDict_GetItemString(py_builtins, "compile"); raise_if_failed(env, py_compile); auto py_module_body = PyObject_GetAttrString(py_module_ast, "body"); raise_if_failed(env, py_module_body); auto py_module_body_guard = PyDecRefGuard(py_module_body); auto py_module_body_size = PyList_Size(py_module_body); raise_if_failed(env, py_module_body_size); if (py_module_body_size > 0) { auto py_last_expr = PyList_GetItem(py_module_body, py_module_body_size - 1); raise_if_failed(env, py_last_expr); auto py_Expr = PyObject_GetAttrString(py_ast, "Expr"); raise_if_failed(env, py_Expr); auto py_Expr_guard = PyDecRefGuard(py_Expr); auto is_Expr_instance = PyObject_IsInstance(py_last_expr, py_Expr); raise_if_failed(env, is_Expr_instance); if (is_Expr_instance) { auto py_module_body_pop = PyObject_GetAttrString(py_module_body, "pop"); raise_if_failed(env, py_module_body_pop); auto py_module_body_pop_guard = PyDecRefGuard(py_module_body_pop); py_module_body_size -= 1; py_last_expr = PyObject_CallNoArgs(py_module_body_pop); raise_if_failed(env, py_last_expr); auto py_last_statement_guard = PyDecRefGuard(py_last_expr); auto py_last_expr_value = PyObject_GetAttrString(py_last_expr, "value"); raise_if_failed(env, py_last_expr_value); auto py_last_expr_value_guard = PyDecRefGuard(py_last_expr_value); auto py_Expression = PyObject_GetAttrString(py_ast, "Expression"); raise_if_failed(env, py_Expression); auto py_Expression_guard = PyDecRefGuard(py_Expression); auto py_Expression_args = PyTuple_Pack(1, py_last_expr_value); raise_if_failed(env, py_Expression_args); auto py_Expression_args_guard = PyDecRefGuard(py_Expression_args); auto py_expr = PyObject_Call(py_Expression, py_Expression_args, NULL); raise_if_failed(env, py_expr); auto py_expr_guard = PyDecRefGuard(py_expr); for (const auto &attr_name : {"lineno", "col_offset", "end_col_offset"}) { auto attr_value = PyObject_GetAttrString(py_last_expr, attr_name); raise_if_failed(env, attr_value); auto attr_value_guard = PyDecRefGuard(attr_value); raise_if_failed(env, PyObject_SetAttrString(py_expr, attr_name, attr_value)); } auto py_eval_string = PyUnicode_FromStringAndSize("eval", 4); raise_if_failed(env, py_eval_string); auto py_eval_string_guard = PyDecRefGuard(py_eval_string); auto py_compile_args = PyTuple_Pack(3, py_expr, py_file_string, py_eval_string); raise_if_failed(env, py_compile_args); auto py_compile_args_guard = PyDecRefGuard(py_compile_args); py_last_expr_code = PyObject_Call(py_compile, py_compile_args, NULL); raise_if_failed(env, py_last_expr_code); py_last_expr_code_guard = py_last_expr_code; } } if (py_module_body_size > 0) { auto py_compile_args = PyTuple_Pack(3, py_module_ast, py_file_string, py_exec_string); raise_if_failed(env, py_compile_args); auto py_compile_args_guard = PyDecRefGuard(py_compile_args); py_body_code = PyObject_Call(py_compile, py_compile_args, NULL); raise_if_failed(env, py_body_code); } py_last_expr_code_guard = nullptr; return std::make_tuple(py_body_code, py_last_expr_code); } std::tuple eval(ErlNifEnv *env, ErlNifBinary code, std::string code_md5, std::vector globals, ErlNifPid stdout_device, ErlNifPid stderr_device) { ensure_initialized(); // Step 1: compile (or get cached result) PyObjectPtr py_body_code = nullptr; PyObjectPtr py_last_expr_code = nullptr; { // Note that it is important that we don't hold GIL while trying // to acquire the mutex, otherwise we could deadlock. auto guard = std::lock_guard(compilation_cache_mutex); if (compilation_cache.find(code_md5) == compilation_cache.end()) { auto gil_guard = PyGILGuard(); auto compiled = compile(env, code); compilation_cache[code_md5] = compiled; } auto compiled = compilation_cache[code_md5]; py_body_code = std::get(compiled); py_last_expr_code = std::get(compiled); } auto gil_guard = PyGILGuard(); // Step 2: prepare globals // For globals, we create a new module named __main__ and use its // dict as globals (extended with the given entries). It corresponds // to the following Python code: // // import types // import sys // // main_module = types.ModuleType("__main__") // main_module.__dict__["__builtins__"] = builtins // sys.modules["__main__"] = main_module // // main_module.__dict__["x"] = 1 // main_module.__dict__["y"] = 2 // // eval(..., main_module.__dict__) // // Note that a more straightforward approach would be to get the // default __main__ module, copy its dict and use that as globals. // However, to better mirror actual Python execution, we want to // use an actual module dict as globals. // // A practical scenario where this matters is pickling an object // of a class defined via evaluation. The pickle module consults // sys.modules (in this case sys.modules["__main__"]) and looks up // the class or function name. If we use a plain dict as globals, // the class will be defined only in that dict and such lookups wil // fail. // // However, it is worth noting that the current approach is not // perfect as it can fail under race conditions. Evaluations may // happen concurrently (if one of them releases GIL, for example, // by calling time.sleep) and there can be other threads started // by evaluation. If a new evaluation sets sys.modules["__main__"] // and yields, an older evaluation or thread may resume and at that // point the value of sys.modules["__main__"] is no longer accurate. // For more details see [1]. // // [1]: https://github.com/marimo-team/marimo/pull/811 auto py_types = PyImport_ImportModule("types"); raise_if_failed(env, py_types); auto py_types_guard = PyDecRefGuard(py_types); auto py_ModuleType = PyObject_GetAttrString(py_types, "ModuleType"); raise_if_failed(env, py_ModuleType); auto py_ModuleType_guard = PyDecRefGuard(py_ModuleType); auto py_main_module_string = PyUnicode_FromStringAndSize("__main__", 8); raise_if_failed(env, py_main_module_string); auto py_main_module_string_guard = PyDecRefGuard(py_main_module_string); auto py_ModuleType_args = PyTuple_Pack(1, py_main_module_string); raise_if_failed(env, py_ModuleType_args); auto py_ModuleType_args_guard = PyDecRefGuard(py_ModuleType_args); auto py_main_module = PyObject_Call(py_ModuleType, py_ModuleType_args, NULL); raise_if_failed(env, py_main_module); auto py_main_module_guard = PyDecRefGuard(py_main_module); auto py_globals = PyModule_GetDict(py_main_module); raise_if_failed(env, py_globals); auto py_builtins = PyEval_GetBuiltins(); raise_if_failed(env, py_builtins); raise_if_failed( env, PyDict_SetItemString(py_globals, "__builtins__", py_builtins)); // The IO capture consists of the following steps: // // 1. We dump EvalInfo into Python bytes and store it in globals // as __pythonx_eval_info_bytes__. // // 2. When IO happens, our custom sys.stdout.write (overriden on // init) retrieves the info from globals and calls the // pythonx_handle_io_write C function, passing the info. // // 3. The pythonx_handle_io_write C function loads EvalInfo and // uses it to send messages to the specified process. // // Each step has a bit more specifics, but they are explained alongside // the corresponding code. // Here we dump the EvalInfo struct into Python bytes object and // put it in globals. We need to pass all of the data, rather than // a pointer, because it may be used during IO from Python threads, // even after the NIF finished. The ErlNifPid struct is opaque, so // we need to store struct memory contents anyway, hence we do this // for the whole EvalInfo struct at once. // // Note that copying struct memory contents into a byte buffer and // vice versa is safe, as long as the struct is always allocated // using its own type (to guarantee proper alignment). Contrarily, // casting a buffer (such as char*) to struct pointer is not safe. auto eval_info = EvalInfo{}; eval_info.stdout_device = stdout_device; eval_info.stderr_device = stderr_device; eval_info.env = env; eval_info.thread_id = std::this_thread::get_id(); auto py_eval_info_bytes = PyBytes_FromStringAndSize( reinterpret_cast(&eval_info), sizeof(EvalInfo)); raise_if_failed(env, py_eval_info_bytes); raise_if_failed(env, PyDict_SetItemString(py_globals, "__pythonx_eval_info_bytes__", py_eval_info_bytes)); auto py_sys = PyImport_AddModule("sys"); raise_if_failed(env, py_sys); auto py_modules = PyObject_GetAttrString(py_sys, "modules"); raise_if_failed(env, py_modules); auto py_modules_guard = PyDecRefGuard(py_modules); raise_if_failed(env, PyDict_SetItemString(py_modules, "__main__", py_main_module)); auto py_globals_initial = PyDict_Copy(py_globals); raise_if_failed(env, py_globals_initial); auto py_globals_guard = PyDecRefGuard(py_globals_initial); for (const auto &[key, value] : globals) { auto py_key = PyUnicode_FromStringAndSize( reinterpret_cast(key.data), key.size); raise_if_failed(env, py_key); auto result = PyDict_SetItem(py_globals, py_key, value.resource->py_object); Py_DecRef(py_key); raise_if_failed(env, result); } // Step 3: eval body and expression if (py_body_code != nullptr) { auto py_body_result = PyEval_EvalCode(py_body_code, py_globals, py_globals); raise_if_failed(env, py_body_result); Py_DecRef(py_body_result); } auto result = std::optional(); if (py_last_expr_code != nullptr) { auto py_result = PyEval_EvalCode(py_last_expr_code, py_globals, py_globals); raise_if_failed(env, py_result); result = ExObject(fine::make_resource(py_result)); } // Step 4: flat-decode globals std::vector key_terms; std::vector value_terms; PyObjectPtr py_key, py_value; Py_ssize_t pos = 0; auto py_str_type = PyDict_GetItemString(py_builtins, "str"); raise_if_failed(env, py_str_type); while (PyDict_Next(py_globals, &pos, &py_key, &py_value)) { // If the key was present in the default globals, ignore it if (PyDict_GetItem(py_globals_initial, py_key) != NULL) { continue; } auto is_unicode = PyObject_IsInstance(py_key, py_str_type); raise_if_failed(env, is_unicode); // If the key is not a string, ignore it. This can happen if the // globals dict is modified directly, but that's not expected. if (!is_unicode) { continue; } auto key_term = py_object_to_binary_term(env, py_key); key_terms.push_back(key_term); // Incref before making the resource Py_IncRef(py_value); auto ex_value = ExObject(fine::make_resource(py_value)); value_terms.push_back(fine::encode(env, ex_value)); } ERL_NIF_TERM map; if (!enif_make_map_from_arrays(env, key_terms.data(), value_terms.data(), key_terms.size(), &map)) { throw std::runtime_error("failed to make a map"); } return std::make_tuple(result, map); } FINE_NIF(eval, ERL_NIF_DIRTY_JOB_CPU_BOUND); } // namespace pythonx FINE_INIT("Elixir.Pythonx.NIF"); // Below are functions we call from Python code extern "C" void pythonx_handle_io_write(const char *message, const char *eval_info_bytes, bool type) { // Note that we allocate EvalInfo first, so it will have the proper // alignment and memcpy simply restores the original struct state. auto eval_info = pythonx::EvalInfo{}; std::memcpy(&eval_info, eval_info_bytes, sizeof(pythonx::EvalInfo)); auto env = enif_alloc_env(); // The enif_whereis_pid and enif_send functions require passing the // caller env. Stdout write may be called by the evaluated code from // the NIF call, but it may also be called by a Python thread, after // the NIF call already finished. Since the BEAM uses OS threads for // its schedulers, we can simply check if this function is invoked // in the same thread as the NIF, or in a different one (Python thread). bool is_main_thread = std::this_thread::get_id() == eval_info.thread_id; auto caller_env = is_main_thread ? eval_info.env : NULL; // Note that we send the output to Pythonx.Janitor and it then sends // it to the device. We do this to avoid IO replies being sent to // the calling Elixir process (which would be unexpected). auto janitor_name = fine::encode(env, pythonx::atoms::ElixirPythonxJanitor); ErlNifPid janitor_pid; if (enif_whereis_pid(caller_env, janitor_name, &janitor_pid)) { auto device = type == 0 ? eval_info.stdout_device : eval_info.stderr_device; auto msg = fine::encode(env, std::make_tuple(pythonx::atoms::output, std::string(message), device)); enif_send(caller_env, &janitor_pid, env, msg); enif_free_env(env); } else { std::cerr

Back | FazBrowse Home | New Git URL