/* Thread and interpreter state structures and their interfaces */

#include "Python.h"

#include "pycore_abstract.h"      // _PyIndex_Check()

#include "pycore_audit.h"         // _Py_AuditHookEntry

#include "pycore_ceval.h"         // _PyEval_AcquireLock()

#include "pycore_codecs.h"        // _PyCodec_Fini()

#include "pycore_critical_section.h" // _PyCriticalSection_Resume()

#include "pycore_dtoa.h"          // _dtoa_state_INIT()

#include "pycore_emscripten_trampoline.h" // _Py_EmscriptenTrampoline_Init()

#include "pycore_freelist.h"      // _PyObject_ClearFreeLists()

#include "pycore_initconfig.h"    // _PyStatus_OK()

#include "pycore_interpframe.h"   // _PyThreadState_HasStackSpace()

#include "pycore_object.h"        // _PyType_InitCache()

#include "pycore_obmalloc.h"      // _PyMem_obmalloc_state_on_heap()

#include "pycore_optimizer.h"     // JIT_CLEANUP_THRESHOLD

#include "pycore_parking_lot.h"   // _PyParkingLot_AfterFork()

#include "pycore_pyerrors.h"      // _PyErr_Clear()

#include "pycore_pylifecycle.h"   // _PyAST_Fini()

#include "pycore_pymem.h"         // _PyMem_DebugEnabled()

#include "pycore_runtime.h"       // _PyRuntime

#include "pycore_runtime_init.h"  // _PyRuntimeState_INIT

#include "pycore_stackref.h"      // Py_STACKREF_DEBUG

#include "pycore_time.h"          // _PyTime_Init()

#include "pycore_uniqueid.h"      // _PyObject_FinalizePerThreadRefcounts()

/* --------------------------------------------------------------------------

CAUTION

Always use PyMem_RawMalloc() and PyMem_RawFree() directly in this file.  A

number of these functions are advertised as safe to call when the GIL isn't

held, and in a debug build Python redirects (e.g.) PyMem_NEW (etc) to Python's

debugging obmalloc functions.  Those aren't thread-safe (they rely on the GIL

to avoid the expense of doing their own locking).

-------------------------------------------------------------------------- */

#ifdef HAVE_DLOPEN

#  ifdef HAVE_DLFCN_H

#    include <dlfcn.h>

#  endif

#  if !HAVE_DECL_RTLD_LAZY

#    define RTLD_LAZY 1

#  endif

#endif

/****************************************/

/* helpers for the current thread state */

/****************************************/

// API for the current thread state is further down.

/* "current" means one of:

   - bound to the current OS thread

   - holds the GIL

 */

//-------------------------------------------------

// a highly efficient lookup for the current thread

//-------------------------------------------------

/*

   The stored thread state is set by PyThreadState_Swap().

   For each of these functions, the GIL must be held by the current thread.

 */

#ifdef HAVE_THREAD_LOCAL

/* The attached thread state for the current thread. */

_Py_thread_local PyThreadState *_Py_tss_tstate = NULL;

/* The "bound" thread state used by PyGILState_Ensure(),

   also known as a "gilstate." */

_Py_thread_local PyThreadState *_Py_tss_gilstate = NULL;

#endif

static inline PyThreadState *

current_fast_get(void)

{

#ifdef HAVE_THREAD_LOCAL

    return _Py_tss_tstate;

#else

    // XXX Fall back to the PyThread_tss_*() API.

#  error "no supported thread-local variable storage classifier"

#endif

}

static inline void

current_fast_set(_PyRuntimeState *Py_UNUSED(runtime), PyThreadState *tstate)

{

    assert(tstate != NULL);

#ifdef HAVE_THREAD_LOCAL

    _Py_tss_tstate = tstate;

#else

    // XXX Fall back to the PyThread_tss_*() API.

#  error "no supported thread-local variable storage classifier"

#endif

}

static inline void

current_fast_clear(_PyRuntimeState *Py_UNUSED(runtime))

{

#ifdef HAVE_THREAD_LOCAL

    _Py_tss_tstate = NULL;

#else

    // XXX Fall back to the PyThread_tss_*() API.

#  error "no supported thread-local variable storage classifier"

#endif

}

#define tstate_verify_not_active(tstate) \

    if (tstate == current_fast_get()) { \

        _Py_FatalErrorFormat(__func__, "tstate %p is still current", tstate); \

    }

PyThreadState *

_PyThreadState_GetCurrent(void)

{

    return current_fast_get();

}

//---------------------------------------------

// The thread state used by PyGILState_Ensure()

//---------------------------------------------

/*

   The stored thread state is set by bind_tstate() (AKA PyThreadState_Bind().

   The GIL does no need to be held for these.

  */

static inline PyThreadState *

gilstate_get(void)

{

    return _Py_tss_gilstate;

}

static inline void

gilstate_set(PyThreadState *tstate)

{

    assert(tstate != NULL);

    _Py_tss_gilstate = tstate;

}

static inline void

gilstate_clear(void)

{

    _Py_tss_gilstate = NULL;

}

#ifndef NDEBUG

static inline int tstate_is_alive(PyThreadState *tstate);

static inline int

tstate_is_bound(PyThreadState *tstate)

{

    return tstate->_status.bound && !tstate->_status.unbound;

}

#endif  // !NDEBUG

static void bind_gilstate_tstate(PyThreadState *);

static void unbind_gilstate_tstate(PyThreadState *);

static void tstate_mimalloc_bind(PyThreadState *);

static void

bind_tstate(PyThreadState *tstate)

{

    assert(tstate != NULL);

    assert(tstate_is_alive(tstate) && !tstate->_status.bound);

    assert(!tstate->_status.unbound);  // just in case

    assert(!tstate->_status.bound_gilstate);

    assert(tstate != gilstate_get());

    assert(!tstate->_status.active);

    assert(tstate->thread_id == 0);

    assert(tstate->native_thread_id == 0);

    // Currently we don't necessarily store the thread state

    // in thread-local storage (e.g. per-interpreter).

    tstate->thread_id = PyThread_get_thread_ident();

#ifdef PY_HAVE_THREAD_NATIVE_ID

    tstate->native_thread_id = PyThread_get_thread_native_id();

#endif

#ifdef Py_GIL_DISABLED

    // Initialize biased reference counting inter-thread queue. Note that this

    // needs to be initialized from the active thread.

    _Py_brc_init_thread(tstate);

#endif

    // mimalloc state needs to be initialized from the active thread.

    tstate_mimalloc_bind(tstate);

    tstate->_status.bound = 1;

}

static void

unbind_tstate(PyThreadState *tstate)

{

    assert(tstate != NULL);

    assert(tstate_is_bound(tstate));

#ifndef HAVE_PTHREAD_STUBS

    assert(tstate->thread_id > 0);

#endif

#ifdef PY_HAVE_THREAD_NATIVE_ID

    assert(tstate->native_thread_id > 0);

#endif

    // We leave thread_id and native_thread_id alone

    // since they can be useful for debugging.

    // Check the `_status` field to know if these values

    // are still valid.

    // We leave tstate->_status.bound set to 1

    // to indicate it was previously bound.

    tstate->_status.unbound = 1;

}

/* Stick the thread state for this thread in thread specific storage.

   When a thread state is created for a thread by some mechanism

   other than PyGILState_Ensure(), it's important that the GILState

   machinery knows about it so it doesn't try to create another

   thread state for the thread.

   (This is a better fix for SF bug #1010677 than the first one attempted.)

   The only situation where you can legitimately have more than one

   thread state for an OS level thread is when there are multiple

   interpreters.

   Before 3.12, the PyGILState_*() APIs didn't work with multiple

   interpreters (see bpo-10915 and bpo-15751), so this function used

   to set TSS only once.  Thus, the first thread state created for that

   given OS level thread would "win", which seemed reasonable behaviour.

*/

static void

bind_gilstate_tstate(PyThreadState *tstate)

{

    assert(tstate != NULL);

    assert(tstate_is_alive(tstate));

    assert(tstate_is_bound(tstate));

    // XXX assert(!tstate->_status.active);

    assert(!tstate->_status.bound_gilstate);

    PyThreadState *tcur = gilstate_get();

    assert(tstate != tcur);

    if (tcur != NULL) {

        tcur->_status.bound_gilstate = 0;

    }

    gilstate_set(tstate);

    tstate->_status.bound_gilstate = 1;

}

static void

unbind_gilstate_tstate(PyThreadState *tstate)

{

    assert(tstate != NULL);

    // XXX assert(tstate_is_alive(tstate));

    assert(tstate_is_bound(tstate));

    // XXX assert(!tstate->_status.active);

    assert(tstate->_status.bound_gilstate);

    assert(tstate == gilstate_get());

    gilstate_clear();

    tstate->_status.bound_gilstate = 0;

}

//----------------------------------------------

// the thread state that currently holds the GIL

//----------------------------------------------

/* This is not exported, as it is not reliable!  It can only

   ever be compared to the state for the *current* thread.

   * If not equal, then it doesn't matter that the actual

     value may change immediately after comparison, as it can't

     possibly change to the current thread's state.

   * If equal, then the current thread holds the lock, so the value can't

     change until we yield the lock.

*/

static int

holds_gil(PyThreadState *tstate)

{

    // XXX Fall back to tstate->interp->runtime->ceval.gil.last_holder

    // (and tstate->interp->runtime->ceval.gil.locked).

    assert(tstate != NULL);

    /* Must be the tstate for this thread */

    assert(tstate == gilstate_get());

    return tstate == current_fast_get();

}

/****************************/

/* the global runtime state */

/****************************/

//----------

// lifecycle

//----------

/* Suppress deprecation warning for PyBytesObject.ob_shash */

_Py_COMP_DIAG_PUSH

_Py_COMP_DIAG_IGNORE_DEPR_DECLS

/* We use "initial" if the runtime gets re-used

   (e.g. Py_Finalize() followed by Py_Initialize().

   Note that we initialize "initial" relative to _PyRuntime,

   to ensure pre-initialized pointers point to the active

   runtime state (and not "initial"). */

static const _PyRuntimeState initial = _PyRuntimeState_INIT(_PyRuntime, "");

_Py_COMP_DIAG_POP

#define LOCKS_INIT(runtime) \

    { \

        &(runtime)->interpreters.mutex, \

        &(runtime)->xi.data_lookup.registry.mutex, \

        &(runtime)->unicode_state.ids.mutex, \

        &(runtime)->imports.extensions.mutex, \

        &(runtime)->ceval.pending_mainthread.mutex, \

        &(runtime)->atexit.mutex, \

        &(runtime)->audit_hooks.mutex, \

        &(runtime)->allocators.mutex, \

        &(runtime)->_main_interpreter.types.mutex, \

        &(runtime)->_main_interpreter.code_state.mutex, \

    }

static void

init_runtime(_PyRuntimeState *runtime,

             void *open_code_hook, void *open_code_userdata,

             _Py_AuditHookEntry *audit_hook_head,

             Py_ssize_t unicode_next_index)

{

    assert(!runtime->preinitializing);

    assert(!runtime->preinitialized);

    assert(!runtime->core_initialized);

    assert(!runtime->initialized);

    assert(!runtime->_initialized);

    runtime->open_code_hook = open_code_hook;

    runtime->open_code_userdata = open_code_userdata;

    runtime->audit_hooks.head = audit_hook_head;

    PyPreConfig_InitPythonConfig(&runtime->preconfig);

    // Set it to the ID of the main thread of the main interpreter.

    runtime->main_thread = PyThread_get_thread_ident();

    runtime->unicode_state.ids.next_index = unicode_next_index;

#if defined(__EMSCRIPTEN__) && defined(PY_CALL_TRAMPOLINE)

    _Py_EmscriptenTrampoline_Init(runtime);

#endif

    runtime->_initialized = 1;

}

PyStatus

_PyRuntimeState_Init(_PyRuntimeState *runtime)

{

    /* We preserve the hook across init, because there is

       currently no public API to set it between runtime

       initialization and interpreter initialization. */

    void *open_code_hook = runtime->open_code_hook;

    void *open_code_userdata = runtime->open_code_userdata;

    _Py_AuditHookEntry *audit_hook_head = runtime->audit_hooks.head;

    // bpo-42882: Preserve next_index value if Py_Initialize()/Py_Finalize()

    // is called multiple times.

    Py_ssize_t unicode_next_index = runtime->unicode_state.ids.next_index;

    if (runtime->_initialized) {

        // Py_Initialize() must be running again.

        // Reset to _PyRuntimeState_INIT.

        memcpy(runtime, &initial, sizeof(*runtime));

        // Preserve the cookie from the original runtime.

        memcpy(runtime->debug_offsets.cookie, _Py_Debug_Cookie, 8);

        assert(!runtime->_initialized);

    }

    PyStatus status = _PyTime_Init(&runtime->time);

    if (_PyStatus_EXCEPTION(status)) {

        return status;

    }

    init_runtime(runtime, open_code_hook, open_code_userdata, audit_hook_head,

                 unicode_next_index);

    return _PyStatus_OK();

}

void

_PyRuntimeState_Fini(_PyRuntimeState *runtime)

{

#ifdef Py_REF_DEBUG

    /* The count is cleared by _Py_FinalizeRefTotal(). */

    assert(runtime->object_state.interpreter_leaks == 0);

#endif

    gilstate_clear();

}

#ifdef HAVE_FORK

/* This function is called from PyOS_AfterFork_Child to ensure that

   newly created child processes do not share locks with the parent. */

PyStatus

_PyRuntimeState_ReInitThreads(_PyRuntimeState *runtime)

{

    // This was initially set in _PyRuntimeState_Init().

    runtime->main_thread = PyThread_get_thread_ident();

    // Clears the parking lot. Any waiting threads are dead. This must be

    // called before releasing any locks that use the parking lot.

    _PyParkingLot_AfterFork();

    // Re-initialize global locks

    PyMutex *locks[] = LOCKS_INIT(runtime);

    for (size_t i = 0; i < Py_ARRAY_LENGTH(locks); i++) {

        _PyMutex_at_fork_reinit(locks[i]);

    }

#ifdef Py_GIL_DISABLED

    for (PyInterpreterState *interp = runtime->interpreters.head;

         interp != NULL; interp = interp->next)

    {

        for (int i = 0; i < NUM_WEAKREF_LIST_LOCKS; i++) {

            _PyMutex_at_fork_reinit(&interp->weakref_locks[i]);

        }

    }

#endif

    _PyTypes_AfterFork();

    _PyThread_AfterFork(&runtime->threads);

    return _PyStatus_OK();

}

#endif

/*************************************/

/* the per-interpreter runtime state */

/*************************************/

//----------

// lifecycle

//----------

/* Calling this indicates that the runtime is ready to create interpreters. */

PyStatus

_PyInterpreterState_Enable(_PyRuntimeState *runtime)

{

    struct pyinterpreters *interpreters = &runtime->interpreters;

    interpreters->next_id = 0;

    return _PyStatus_OK();

}

static PyInterpreterState *

alloc_interpreter(void)

{

    size_t alignment = _Alignof(PyInterpreterState);

    size_t allocsize = sizeof(PyInterpreterState) + alignment - 1;

    void *mem = PyMem_RawCalloc(1, allocsize);

    if (mem == NULL) {

        return NULL;

    }

    PyInterpreterState *interp = _Py_ALIGN_UP(mem, alignment);

    assert(_Py_IS_ALIGNED(interp, alignment));

    interp->_malloced = mem;

    return interp;

}

static void

free_interpreter(PyInterpreterState *interp)

{

    // The main interpreter is statically allocated so

    // should not be freed.

    if (interp != &_PyRuntime._main_interpreter) {

        if (_PyMem_obmalloc_state_on_heap(interp)) {

            // interpreter has its own obmalloc state, free it

            PyMem_RawFree(interp->obmalloc);

            interp->obmalloc = NULL;

        }

        assert(_Py_IS_ALIGNED(interp, _Alignof(PyInterpreterState)));

        PyMem_RawFree(interp->_malloced);

    }

}

#ifndef NDEBUG

static inline int check_interpreter_whence(long);

#endif

extern _Py_CODEUNIT *

_Py_LazyJitTrampoline(

    struct _PyExecutorObject *exec, _PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate

);

/* Get the interpreter state to a minimal consistent state.

   Further init happens in pylifecycle.c before it can be used.

   All fields not initialized here are expected to be zeroed out,

   e.g. by PyMem_RawCalloc() or memset(), or otherwise pre-initialized.

   The runtime state is not manipulated.  Instead it is assumed that

   the interpreter is getting added to the runtime.

   Note that the main interpreter was statically initialized as part

   of the runtime and most state is already set properly.  That leaves

   a small number of fields to initialize dynamically, as well as some

   that are initialized lazily.

   For subinterpreters we memcpy() the main interpreter in

   PyInterpreterState_New(), leaving it in the same mostly-initialized

   state.  The only difference is that the interpreter has some

   self-referential state that is statically initializexd to the

   main interpreter.  We fix those fields here, in addition

   to the other dynamically initialized fields.

  */

static PyStatus

init_interpreter(PyInterpreterState *interp,

                 _PyRuntimeState *runtime, int64_t id,

                 PyInterpreterState *next,

                 long whence)

{

    if (interp->_initialized) {

        return _PyStatus_ERR("interpreter already initialized");

    }

    assert(interp->_whence == _PyInterpreterState_WHENCE_NOTSET);

    assert(check_interpreter_whence(whence) == 0);

    interp->_whence = whence;

    assert(runtime != NULL);

    interp->runtime = runtime;

    assert(id > 0 || (id == 0 && interp == runtime->interpreters.main));

    interp->id = id;

    interp->id_refcount = 0;

    assert(runtime->interpreters.head == interp);

    assert(next != NULL || (interp == runtime->interpreters.main));

    interp->next = next;

    interp->threads.preallocated = &interp->_initial_thread;

    // We would call _PyObject_InitState() at this point

    // if interp->feature_flags were alredy set.

    _PyEval_InitState(interp);

    _PyGC_InitState(&interp->gc);

    PyConfig_InitPythonConfig(&interp->config);

    _PyType_InitCache(interp);

#ifdef Py_GIL_DISABLED

    _Py_brc_init_state(interp);

#endif

    llist_init(&interp->mem_free_queue.head);

    llist_init(&interp->asyncio_tasks_head);

    interp->asyncio_tasks_lock = (PyMutex){0};

    for (int i = 0; i < _PY_MONITORING_UNGROUPED_EVENTS; i++) {

        interp->monitors.tools[i] = 0;

    }

    for (int t = 0; t < PY_MONITORING_TOOL_IDS; t++) {

        for (int e = 0; e < _PY_MONITORING_EVENTS; e++) {

            interp->monitoring_callables[t][e] = NULL;

        }

        interp->monitoring_tool_versions[t] = 0;

    }

    interp->_code_object_generation = 0;

    interp->jit = false;

    interp->executor_list_head = NULL;

    interp->executor_deletion_list_head = NULL;

    interp->executor_deletion_list_remaining_capacity = 0;

    interp->trace_run_counter = JIT_CLEANUP_THRESHOLD;

    if (interp != &runtime->_main_interpreter) {

        /* Fix the self-referential, statically initialized fields. */

        interp->dtoa = (struct _dtoa_state)_dtoa_state_INIT(interp);

    }

#if !defined(Py_GIL_DISABLED) && defined(Py_STACKREF_DEBUG)

    interp->next_stackref = INITIAL_STACKREF_INDEX;

    _Py_hashtable_allocator_t alloc = {

        .malloc = malloc,

        .free = free,

    };

    interp->open_stackrefs_table = _Py_hashtable_new_full(

        _Py_hashtable_hash_ptr,

        _Py_hashtable_compare_direct,

        NULL,

        NULL,

        &alloc

    );

#  ifdef Py_STACKREF_CLOSE_DEBUG

    interp->closed_stackrefs_table = _Py_hashtable_new_full(

        _Py_hashtable_hash_ptr,

        _Py_hashtable_compare_direct,

        NULL,

        NULL,

        &alloc

    );

#  endif

    _Py_stackref_associate(interp, Py_None, PyStackRef_None);

    _Py_stackref_associate(interp, Py_False, PyStackRef_False);

    _Py_stackref_associate(interp, Py_True, PyStackRef_True);

#endif

    interp->_initialized = 1;

    return _PyStatus_OK();

}

PyStatus

_PyInterpreterState_New(PyThreadState *tstate, PyInterpreterState **pinterp)

{

    *pinterp = NULL;

    // Don't get runtime from tstate since tstate can be NULL

    _PyRuntimeState *runtime = &_PyRuntime;

    // tstate is NULL when pycore_create_interpreter() calls

    // _PyInterpreterState_New() to create the main interpreter.

    if (tstate != NULL) {

        if (_PySys_Audit(tstate, "cpython.PyInterpreterState_New", NULL) < 0) {

            return _PyStatus_ERR("sys.audit failed");

        }

    }

    /* We completely serialize creation of multiple interpreters, since

       it simplifies things here and blocking concurrent calls isn't a problem.

       Regardless, we must fully block subinterpreter creation until

       after the main interpreter is created. */

    HEAD_LOCK(runtime);

    struct pyinterpreters *interpreters = &runtime->interpreters;

    int64_t id = interpreters->next_id;

    interpreters->next_id += 1;

    // Allocate the interpreter and add it to the runtime state.

    PyInterpreterState *interp;

    PyStatus status;

    PyInterpreterState *old_head = interpreters->head;

    if (old_head == NULL) {

        // We are creating the main interpreter.

        assert(interpreters->main == NULL);

        assert(id == 0);

        interp = &runtime->_main_interpreter;

        assert(interp->id == 0);

        assert(interp->next == NULL);

        interpreters->main = interp;

    }

    else {

        assert(interpreters->main != NULL);

        assert(id != 0);

        interp = alloc_interpreter();

        if (interp == NULL) {

            status = _PyStatus_NO_MEMORY();

            goto error;

        }

        // Set to _PyInterpreterState_INIT.

        memcpy(interp, &initial._main_interpreter, sizeof(*interp));

        if (id < 0) {

            /* overflow or Py_Initialize() not called yet! */

            status = _PyStatus_ERR("failed to get an interpreter ID");

            goto error;

        }

    }

    interpreters->head = interp;

    long whence = _PyInterpreterState_WHENCE_UNKNOWN;

    status = init_interpreter(interp, runtime,

                              id, old_head, whence);

    if (_PyStatus_EXCEPTION(status)) {

        goto error;

    }

    HEAD_UNLOCK(runtime);

    assert(interp != NULL);

    *pinterp = interp;

    return _PyStatus_OK();

error:

    HEAD_UNLOCK(runtime);

    if (interp != NULL) {

        free_interpreter(interp);

    }

    return status;

}

PyInterpreterState *

PyInterpreterState_New(void)

{

    // tstate can be NULL

    PyThreadState *tstate = current_fast_get();

    PyInterpreterState *interp;

    PyStatus status = _PyInterpreterState_New(tstate, &interp);

    if (_PyStatus_EXCEPTION(status)) {

        Py_ExitStatusException(status);

    }

    assert(interp != NULL);

    return interp;

}

#if !defined(Py_GIL_DISABLED) && defined(Py_STACKREF_DEBUG)

extern void

_Py_stackref_report_leaks(PyInterpreterState *interp);

#endif

static void

interpreter_clear(PyInterpreterState *interp, PyThreadState *tstate)

{

    assert(interp != NULL);

    assert(tstate != NULL);

    _PyRuntimeState *runtime = interp->runtime;

    /* XXX Conditions we need to enforce:

       * the GIL must be held by the current thread

       * tstate must be the "current" thread state (current_fast_get())

       * tstate->interp must be interp

       * for the main interpreter, tstate must be the main thread

     */

    // XXX Ideally, we would not rely on any thread state in this function

    // (and we would drop the "tstate" argument).

    if (_PySys_Audit(tstate, "cpython.PyInterpreterState_Clear", NULL) < 0) {

        _PyErr_Clear(tstate);

    }

    // Clear the current/main thread state last.

    _Py_FOR_EACH_TSTATE_BEGIN(interp, p) {

        // See https://github.com/python/cpython/issues/102126

        // Must be called without HEAD_LOCK held as it can deadlock

        // if any finalizer tries to acquire that lock.

        HEAD_UNLOCK(runtime);

        PyThreadState_Clear(p);

        HEAD_LOCK(runtime);

    }

    _Py_FOR_EACH_TSTATE_END(interp);

    if (tstate->interp == interp) {

        /* We fix tstate->_status below when we for sure aren't using it

           (e.g. no longer need the GIL). */

        // XXX Eliminate the need to do this.

        tstate->_status.cleared = 0;

    }

    /* It is possible that any of the objects below have a finalizer

       that runs Python code or otherwise relies on a thread state

       or even the interpreter state.  For now we trust that isn't

       a problem.

     */

    // XXX Make sure we properly deal with problematic finalizers.

    Py_CLEAR(interp->audit_hooks);

    // At this time, all the threads should be cleared so we don't need atomic

    // operations for instrumentation_version or eval_breaker.

    interp->ceval.instrumentation_version = 0;

    tstate->eval_breaker = 0;

    for (int i = 0; i < _PY_MONITORING_UNGROUPED_EVENTS; i++) {

        interp->monitors.tools[i] = 0;

    }

    for (int t = 0; t < PY_MONITORING_TOOL_IDS; t++) {

        for (int e = 0; e < _PY_MONITORING_EVENTS; e++) {

            Py_CLEAR(interp->monitoring_callables[t][e]);

        }

    }

    for (int t = 0; t < PY_MONITORING_TOOL_IDS; t++) {

        Py_CLEAR(interp->monitoring_tool_names[t]);

    }

    interp->_code_object_generation = 0;

#ifdef Py_GIL_DISABLED

    interp->tlbc_indices.tlbc_generation = 0;

#endif

    PyConfig_Clear(&interp->config);

    _PyCodec_Fini(interp);

    assert(interp->imports.modules == NULL);

    assert(interp->imports.modules_by_index == NULL);

    assert(interp->imports.importlib == NULL);

    assert(interp->imports.import_func == NULL);

    Py_CLEAR(interp->sysdict_copy);

    Py_CLEAR(interp->builtins_copy);

    Py_CLEAR(interp->dict);

#ifdef HAVE_FORK

    Py_CLEAR(interp->before_forkers);

    Py_CLEAR(interp->after_forkers_parent);

    Py_CLEAR(interp->after_forkers_child);

#endif

#ifdef _Py_TIER2

    _Py_ClearExecutorDeletionList(interp);

#endif

    _PyAST_Fini(interp);

    _PyAtExit_Fini(interp);

    // All Python types must be destroyed before the last GC collection. Python

    // types create a reference cycle to themselves in their in their

    // PyTypeObject.tp_mro member (the tuple contains the type).

    /* Last garbage collection on this interpreter */

    _PyGC_CollectNoFail(tstate);

    _PyGC_Fini(interp);

    // Finalize warnings after last gc so that any finalizers can

    // access warnings state

    _PyWarnings_Fini(interp);

    struct _PyExecutorObject *cold = interp->cold_executor;

    if (cold != NULL) {

        interp->cold_executor = NULL;

        assert(cold->vm_data.valid);

        assert(cold->vm_data.warm);

        _PyExecutor_Free(cold);

    }

    /* We don't clear sysdict and builtins until the end of this function.

       Because clearing other attributes can execute arbitrary Python code

       which requires sysdict and builtins. */

    PyDict_Clear(interp->sysdict);

    PyDict_Clear(interp->builtins);

    Py_CLEAR(interp->sysdict);

    Py_CLEAR(interp->builtins);

#if !defined(Py_GIL_DISABLED) && defined(Py_STACKREF_DEBUG)

#  ifdef Py_STACKREF_CLOSE_DEBUG

    _Py_hashtable_destroy(interp->closed_stackrefs_table);

    interp->closed_stackrefs_table = NULL;

#  endif

    _Py_stackref_report_leaks(interp);

    _Py_hashtable_destroy(interp->open_stackrefs_table);

    interp->open_stackrefs_table = NULL;

#endif

    if (tstate->interp == interp) {

        /* We are now safe to fix tstate->_status.cleared. */

        // XXX Do this (much) earlier?

        tstate->_status.cleared = 1;

    }

    for (int i=0; i < DICT_MAX_WATCHERS; i++) {

        interp->dict_state.watchers[i] = NULL;

    }

    for (int i=0; i < TYPE_MAX_WATCHERS; i++) {

        interp->type_watchers[i] = NULL;

    }

    for (int i=0; i < FUNC_MAX_WATCHERS; i++) {

        interp->func_watchers[i] = NULL;

    }

    interp->active_func_watchers = 0;

    for (int i=0; i < CODE_MAX_WATCHERS; i++) {

        interp->code_watchers[i] = NULL;

    }

    interp->active_code_watchers = 0;

    for (int i=0; i < CONTEXT_MAX_WATCHERS; i++) {

        interp->context_watchers[i] = NULL;

    }

    interp->active_context_watchers = 0;

    // XXX Once we have one allocator per interpreter (i.e.

    // per-interpreter GC) we must ensure that all of the interpreter's

    // objects have been cleaned up at the point.

    // We could clear interp->threads.freelist here

    // if it held more than just the initial thread state.

}

void

PyInterpreterState_Clear(PyInterpreterState *interp)

{

    // Use the current Python thread state to call audit hooks and to collect

    // garbage. It can be different than the current Python thread state

    // of 'interp'.

    PyThreadState *current_tstate = current_fast_get();

    _PyImport_ClearCore(interp);

    interpreter_clear(interp, current_tstate);

}

void

_PyInterpreterState_Clear(PyThreadState *tstate)

{

    _PyImport_ClearCore(tstate->interp);

    interpreter_clear(tstate->interp, tstate);

}

static inline void tstate_deactivate(PyThreadState *tstate);

static void tstate_set_detached(PyThreadState *tstate, int detached_state);

static void zapthreads(PyInterpreterState *interp);

void

PyInterpreterState_Delete(PyInterpreterState *interp)

{

    _PyRuntimeState *runtime = interp->runtime;

    struct pyinterpreters *interpreters = &runtime->interpreters;

    // XXX Clearing the "current" thread state should happen before

    // we start finalizing the interpreter (or the current thread state).

    PyThreadState *tcur = current_fast_get();

    if (tcur != NULL && interp == tcur->interp) {

        /* Unset current thread.  After this, many C API calls become crashy. */

        _PyThreadState_Detach(tcur);

    }

    zapthreads(interp);

    // XXX These two calls should be done at the end of clear_interpreter(),

    // but currently some objects get decref'ed after that.

#ifdef Py_REF_DEBUG

    _PyInterpreterState_FinalizeRefTotal(interp);

#endif

    _PyInterpreterState_FinalizeAllocatedBlocks(interp);

    HEAD_LOCK(runtime);

    PyInterpreterState **p;

    for (p = &interpreters->head; ; p = &(*p)->next) {

        if (*p == NULL) {

            Py_FatalError("NULL interpreter");

        }

        if (*p == interp) {

            break;

        }

    }

    if (interp->threads.head != NULL) {

        Py_FatalError("remaining threads");

    }

    *p = interp->next;

    if (interpreters->main == interp) {

        interpreters->main = NULL;

        if (interpreters->head != NULL) {

            Py_FatalError("remaining subinterpreters");

        }

    }

    HEAD_UNLOCK(runtime);

    _Py_qsbr_fini(interp);

    _PyObject_FiniState(interp);

    free_interpreter(interp);

}

#ifdef HAVE_FORK

/*

 * Delete all interpreter states except the main interpreter.  If there

 * is a current interpreter state, it *must* be the main interpreter.

 */

PyStatus

_PyInterpreterState_DeleteExceptMain(_PyRuntimeState *runtime)

{

    struct pyinterpreters *interpreters = &runtime->interpreters;

    PyThreadState *tstate = _PyThreadState_Swap(runtime, NULL);

    if (tstate != NULL && tstate->interp != interpreters->main) {

        return _PyStatus_ERR("not main interpreter");

    }

    HEAD_LOCK(runtime);

    PyInterpreterState *interp = interpreters->head;

    interpreters->head = NULL;

    while (interp != NULL) {

        if (interp == interpreters->main) {

            interpreters->main->next = NULL;

            interpreters->head = interp;

            interp = interp->next;

            continue;

        }

        // XXX Won't this fail since PyInterpreterState_Clear() requires

        // the "current" tstate to be set?

        PyInterpreterState_Clear(interp);  // XXX must activate?

        zapthreads(interp);

        PyInterpreterState *prev_interp = interp;

        interp = interp->next;

        free_interpreter(prev_interp);

    }

    HEAD_UNLOCK(runtime);

    if (interpreters->head == NULL) {

        return _PyStatus_ERR("missing main interpreter");