bpo-33608: Factor out a private, per-interpreter _Py_AddPendingCall(). (GH-11617)

This involves moving the global "pending calls" state to PyInterpreterState.

https://bugs.python.org/issue33608

10 files changed, 201 insertions, 121 deletions

diff --git a/Include/ceval.h b/Include/ceval.h
index 11283c0..9c6d420 100644
--- a/Include/ceval.h
+++ b/Include/ceval.h
@@ -221,7 +221,7 @@ PyAPI_FUNC(Py_ssize_t) _PyEval_RequestCodeExtraIndex(freefunc);
 #ifndef Py_LIMITED_API
 PyAPI_FUNC(int) _PyEval_SliceIndex(PyObject *, Py_ssize_t *);
 PyAPI_FUNC(int) _PyEval_SliceIndexNotNone(PyObject *, Py_ssize_t *);
-PyAPI_FUNC(void) _PyEval_SignalAsyncExc(void);
+PyAPI_FUNC(void) _PyEval_SignalAsyncExc(PyInterpreterState *);
 #endif
 
 /* Masks and values used by FORMAT_VALUE opcode. */
diff --git a/Include/internal/pycore_ceval.h b/Include/internal/pycore_ceval.h
index b9f2d7d..5a80f6f 100644
--- a/Include/internal/pycore_ceval.h
+++ b/Include/internal/pycore_ceval.h
@@ -11,8 +11,12 @@ extern "C" {
 #include "pycore_atomic.h"
 #include "pythread.h"
 
+struct _is;  // See PyInterpreterState in cpython/pystate.h.
+
+PyAPI_FUNC(int) _Py_AddPendingCall(struct _is*, unsigned long, int (*)(void *), void *);
+PyAPI_FUNC(int) _Py_MakePendingCalls(struct _is*);
+
 struct _pending_calls {
-    unsigned long main_thread;
     PyThread_type_lock lock;
     /* Request for running pending calls. */
     _Py_atomic_int calls_to_do;
@@ -22,6 +26,7 @@ struct _pending_calls {
     int async_exc;
 #define NPENDINGCALLS 32
     struct {
+        unsigned long thread_id;
         int (*func)(void *);
         void *arg;
     } calls[NPENDINGCALLS];
@@ -29,6 +34,13 @@ struct _pending_calls {
     int last;
 };
 
+struct _ceval_interpreter_state {
+    /* This single variable consolidates all requests to break out of
+       the fast path in the eval loop. */
+    _Py_atomic_int eval_breaker;
+    struct _pending_calls pending;
+};
+
 #include "pycore_gil.h"
 
 struct _ceval_runtime_state {
@@ -39,12 +51,8 @@ struct _ceval_runtime_state {
        c_tracefunc.  This speeds up the if statement in
        PyEval_EvalFrameEx() after fast_next_opcode. */
     int tracing_possible;
-    /* This single variable consolidates all requests to break out of
-       the fast path in the eval loop. */
-    _Py_atomic_int eval_breaker;
     /* Request for dropping the GIL */
     _Py_atomic_int gil_drop_request;
-    struct _pending_calls pending;
     /* Request for checking signals. */
     _Py_atomic_int signals_pending;
     struct _gil_runtime_state gil;
diff --git a/Include/internal/pycore_pystate.h b/Include/internal/pycore_pystate.h
index 7796223..f6c61e7 100644
--- a/Include/internal/pycore_pystate.h
+++ b/Include/internal/pycore_pystate.h
@@ -11,6 +11,7 @@ extern "C" {
 #include "pystate.h"
 #include "pythread.h"
 
+#include "pycore_atomic.h"
 #include "pycore_ceval.h"
 #include "pycore_pathconfig.h"
 #include "pycore_pymem.h"
@@ -31,6 +32,8 @@ struct _is {
     int64_t id_refcount;
     PyThread_type_lock id_mutex;
 
+    int finalizing;
+
     PyObject *modules;
     PyObject *modules_by_index;
     PyObject *sysdict;
@@ -78,6 +81,8 @@ struct _is {
     PyObject *pyexitmodule;
 
     uint64_t tstate_next_unique_id;
+
+    struct _ceval_interpreter_state ceval;
 };
 
 PyAPI_FUNC(struct _is*) _PyInterpreterState_LookUpID(PY_INT64_T);
@@ -207,6 +212,8 @@ typedef struct pyruntimestate {
         struct _xidregitem *head;
     } xidregistry;
 
+    unsigned long main_thread;
+
 #define NEXITFUNCS 32
     void (*exitfuncs[NEXITFUNCS])(void);
     int nexitfuncs;
diff --git a/Misc/NEWS.d/next/Core and Builtins/2018-09-15-12-13-46.bpo-33608.avmvVP.rst b/Misc/NEWS.d/next/Core and Builtins/2018-09-15-12-13-46.bpo-33608.avmvVP.rst
new file mode 100644
index 0000000..73a01a1
--- /dev/null
+++ b/Misc/NEWS.d/next/Core and Builtins/2018-09-15-12-13-46.bpo-33608.avmvVP.rst
@@ -0,0 +1,5 @@
+We added a new internal _Py_AddPendingCall() that operates relative to the
+provided interpreter.  This allows us to use the existing implementation to
+ask another interpreter to do work that cannot be done in the current
+interpreter, like decref an object the other interpreter owns.  The existing
+Py_AddPendingCall() only operates relative to the main interpreter.
diff --git a/Modules/_testcapimodule.c b/Modules/_testcapimodule.c
index 350ef77..5b0da78 100644
--- a/Modules/_testcapimodule.c
+++ b/Modules/_testcapimodule.c
@@ -2445,6 +2445,7 @@ pending_threadfunc(PyObject *self, PyObject *arg)
     Py_INCREF(callable);
 
     Py_BEGIN_ALLOW_THREADS
+    /* XXX Use the internal _Py_AddPendingCall(). */
     r = Py_AddPendingCall(&_pending_callback, callable);
     Py_END_ALLOW_THREADS
 
diff --git a/Modules/signalmodule.c b/Modules/signalmodule.c
index f29720b..ac3b6c7 100644
--- a/Modules/signalmodule.c
+++ b/Modules/signalmodule.c
@@ -19,6 +19,7 @@
 #include <process.h>
 #endif
 #endif
+#include "internal/pycore_pystate.h"
 
 #ifdef HAVE_SIGNAL_H
 #include <signal.h>
@@ -295,8 +296,10 @@ trip_signal(int sig_num)
                 {
                     /* Py_AddPendingCall() isn't signal-safe, but we
                        still use it for this exceptional case. */
-                    Py_AddPendingCall(report_wakeup_send_error,
-                                      (void *)(intptr_t) last_error);
+                    _Py_AddPendingCall(_PyRuntime.interpreters.main,
+                                       main_thread,
+                                       report_wakeup_send_error,
+                                       (void *)(intptr_t) last_error);
                 }
             }
         }
@@ -313,8 +316,10 @@ trip_signal(int sig_num)
                 {
                     /* Py_AddPendingCall() isn't signal-safe, but we
                        still use it for this exceptional case. */
-                    Py_AddPendingCall(report_wakeup_write_error,
-                                      (void *)(intptr_t)errno);
+                    _Py_AddPendingCall(_PyRuntime.interpreters.main,
+                                       main_thread,
+                                       report_wakeup_write_error,
+                                       (void *)(intptr_t)errno);
                 }
             }
         }
diff --git a/Python/ceval.c b/Python/ceval.c
index 439f4f1..4e139ce 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -96,61 +96,61 @@ static long dxp[256];
 /* This can set eval_breaker to 0 even though gil_drop_request became
    1.  We believe this is all right because the eval loop will release
    the GIL eventually anyway. */
-#define COMPUTE_EVAL_BREAKER() \
+#define COMPUTE_EVAL_BREAKER(interp) \
     _Py_atomic_store_relaxed( \
-        &_PyRuntime.ceval.eval_breaker, \
+        &interp->ceval.eval_breaker, \
         GIL_REQUEST | \
         _Py_atomic_load_relaxed(&_PyRuntime.ceval.signals_pending) | \
-        _Py_atomic_load_relaxed(&_PyRuntime.ceval.pending.calls_to_do) | \
-        _PyRuntime.ceval.pending.async_exc)
+        _Py_atomic_load_relaxed(&interp->ceval.pending.calls_to_do) | \
+        interp->ceval.pending.async_exc)
 
-#define SET_GIL_DROP_REQUEST() \
+#define SET_GIL_DROP_REQUEST(interp) \
     do { \
         _Py_atomic_store_relaxed(&_PyRuntime.ceval.gil_drop_request, 1); \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.eval_breaker, 1); \
+        _Py_atomic_store_relaxed(&interp->ceval.eval_breaker, 1); \
     } while (0)
 
-#define RESET_GIL_DROP_REQUEST() \
+#define RESET_GIL_DROP_REQUEST(interp) \
     do { \
         _Py_atomic_store_relaxed(&_PyRuntime.ceval.gil_drop_request, 0); \
-        COMPUTE_EVAL_BREAKER(); \
+        COMPUTE_EVAL_BREAKER(interp); \
     } while (0)
 
 /* Pending calls are only modified under pending_lock */
-#define SIGNAL_PENDING_CALLS() \
+#define SIGNAL_PENDING_CALLS(interp) \
     do { \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.pending.calls_to_do, 1); \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.eval_breaker, 1); \
+        _Py_atomic_store_relaxed(&interp->ceval.pending.calls_to_do, 1); \
+        _Py_atomic_store_relaxed(&interp->ceval.eval_breaker, 1); \
     } while (0)
 
-#define UNSIGNAL_PENDING_CALLS() \
+#define UNSIGNAL_PENDING_CALLS(interp) \
     do { \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.pending.calls_to_do, 0); \
-        COMPUTE_EVAL_BREAKER(); \
+        _Py_atomic_store_relaxed(&interp->ceval.pending.calls_to_do, 0); \
+        COMPUTE_EVAL_BREAKER(interp); \
     } while (0)
 
 #define SIGNAL_PENDING_SIGNALS() \
     do { \
         _Py_atomic_store_relaxed(&_PyRuntime.ceval.signals_pending, 1); \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.eval_breaker, 1); \
+        _Py_atomic_store_relaxed(&_PyRuntime.interpreters.main->ceval.eval_breaker, 1); \
     } while (0)
 
 #define UNSIGNAL_PENDING_SIGNALS() \
     do { \
         _Py_atomic_store_relaxed(&_PyRuntime.ceval.signals_pending, 0); \
-        COMPUTE_EVAL_BREAKER(); \
+        COMPUTE_EVAL_BREAKER(_PyRuntime.interpreters.main); \
     } while (0)
 
-#define SIGNAL_ASYNC_EXC() \
+#define SIGNAL_ASYNC_EXC(interp) \
     do { \
-        _PyRuntime.ceval.pending.async_exc = 1; \
-        _Py_atomic_store_relaxed(&_PyRuntime.ceval.eval_breaker, 1); \
+        interp->ceval.pending.async_exc = 1; \
+        _Py_atomic_store_relaxed(&interp->ceval.eval_breaker, 1); \
     } while (0)
 
-#define UNSIGNAL_ASYNC_EXC() \
+#define UNSIGNAL_ASYNC_EXC(interp) \
     do { \
-        _PyRuntime.ceval.pending.async_exc = 0; \
-        COMPUTE_EVAL_BREAKER(); \
+        interp->ceval.pending.async_exc = 0; \
+        COMPUTE_EVAL_BREAKER(interp); \
     } while (0)
 
 
@@ -174,9 +174,6 @@ PyEval_InitThreads(void)
     PyThread_init_thread();
     create_gil();
     take_gil(_PyThreadState_GET());
-    _PyRuntime.ceval.pending.main_thread = PyThread_get_thread_ident();
-    if (!_PyRuntime.ceval.pending.lock)
-        _PyRuntime.ceval.pending.lock = PyThread_allocate_lock();
 }
 
 void
@@ -243,9 +240,11 @@ PyEval_ReInitThreads(void)
     if (!gil_created())
         return;
     recreate_gil();
-    _PyRuntime.ceval.pending.lock = PyThread_allocate_lock();
+    // This will be reset in make_pending_calls() below.
+    current_tstate->interp->ceval.pending.lock = NULL;
+
     take_gil(current_tstate);
-    _PyRuntime.ceval.pending.main_thread = PyThread_get_thread_ident();
+    _PyRuntime.main_thread = PyThread_get_thread_ident();
 
     /* Destroy all threads except the current one */
     _PyThreadState_DeleteExcept(current_tstate);
@@ -255,9 +254,9 @@ PyEval_ReInitThreads(void)
    raised. */
 
 void
-_PyEval_SignalAsyncExc(void)
+_PyEval_SignalAsyncExc(PyInterpreterState *interp)
 {
-    SIGNAL_ASYNC_EXC();
+    SIGNAL_ASYNC_EXC(interp);
 }
 
 PyThreadState *
@@ -323,17 +322,58 @@ _PyEval_SignalReceived(void)
     SIGNAL_PENDING_SIGNALS();
 }
 
+static int
+_add_pending_call(PyInterpreterState *interp, unsigned long thread_id, int (*func)(void *), void *arg)
+{
+    int i = interp->ceval.pending.last;
+    int j = (i + 1) % NPENDINGCALLS;
+    if (j == interp->ceval.pending.first) {
+        return -1; /* Queue full */
+    }
+    interp->ceval.pending.calls[i].thread_id = thread_id;
+    interp->ceval.pending.calls[i].func = func;
+    interp->ceval.pending.calls[i].arg = arg;
+    interp->ceval.pending.last = j;
+    return 0;
+}
+
+/* pop one item off the queue while holding the lock */
+static void
+_pop_pending_call(PyInterpreterState *interp, int (**func)(void *), void **arg)
+{
+    int i = interp->ceval.pending.first;
+    if (i == interp->ceval.pending.last) {
+        return; /* Queue empty */
+    }
+
+    *func = interp->ceval.pending.calls[i].func;
+    *arg = interp->ceval.pending.calls[i].arg;
+    interp->ceval.pending.first = (i + 1) % NPENDINGCALLS;
+
+    unsigned long thread_id = interp->ceval.pending.calls[i].thread_id;
+    if (thread_id && PyThread_get_thread_ident() != thread_id) {
+        // Thread mismatch, so move it to the end of the list
+        // and start over.
+        _Py_AddPendingCall(interp, thread_id, *func, *arg);
+        return;
+    }
+}
+
+int
+Py_AddPendingCall(int (*func)(void *), void *arg)
+{
+    PyInterpreterState *interp = _PyRuntime.interpreters.main;
+    return _Py_AddPendingCall(interp, _PyRuntime.main_thread, func, arg);
+}
+
 /* This implementation is thread-safe.  It allows
    scheduling to be made from any thread, and even from an executing
    callback.
  */
 
 int
-Py_AddPendingCall(int (*func)(void *), void *arg)
+_Py_AddPendingCall(PyInterpreterState *interp, unsigned long thread_id, int (*func)(void *), void *arg)
 {
-    int i, j, result=0;
-    PyThread_type_lock lock = _PyRuntime.ceval.pending.lock;
-
     /* try a few times for the lock.  Since this mechanism is used
      * for signal handling (on the main thread), there is a (slim)
      * chance that a signal is delivered on the same thread while we
@@ -345,7 +385,9 @@ Py_AddPendingCall(int (*func)(void *), void *arg)
      * We also check for lock being NULL, in the unlikely case that
      * this function is called before any bytecode evaluation takes place.
      */
+    PyThread_type_lock lock = interp->ceval.pending.lock;
     if (lock != NULL) {
+        int i;
         for (i = 0; i<100; i++) {
             if (PyThread_acquire_lock(lock, NOWAIT_LOCK))
                 break;
@@ -354,17 +396,21 @@ Py_AddPendingCall(int (*func)(void *), void *arg)
             return -1;
     }
 
-    i = _PyRuntime.ceval.pending.last;
-    j = (i + 1) % NPENDINGCALLS;
-    if (j == _PyRuntime.ceval.pending.first) {
-        result = -1; /* Queue full */
-    } else {
-        _PyRuntime.ceval.pending.calls[i].func = func;
-        _PyRuntime.ceval.pending.calls[i].arg = arg;
-        _PyRuntime.ceval.pending.last = j;
+    int result = -1;
+    if (interp->finalizing) {
+        PyObject *exc, *val, *tb;
+        PyErr_Fetch(&exc, &val, &tb);
+        PyErr_SetString(PyExc_SystemError, "Py_AddPendingCall: cannot add pending calls (interpreter shutting down)");
+        PyErr_Print();
+        PyErr_Restore(exc, val, tb);
+        goto done;
     }
+
+    result = _add_pending_call(interp, thread_id, func, arg);
     /* signal main loop */
-    SIGNAL_PENDING_CALLS();
+    SIGNAL_PENDING_CALLS(interp);
+
+done:
     if (lock != NULL)
         PyThread_release_lock(lock);
     return result;
@@ -374,9 +420,7 @@ static int
 handle_signals(void)
 {
     /* Only handle signals on main thread. */
-    if (_PyRuntime.ceval.pending.main_thread &&
-        PyThread_get_thread_ident() != _PyRuntime.ceval.pending.main_thread)
-    {
+    if (PyThread_get_thread_ident() != _PyRuntime.main_thread) {
         return 0;
     }
     /*
@@ -396,17 +440,10 @@ handle_signals(void)
 }
 
 static int
-make_pending_calls(void)
+make_pending_calls(PyInterpreterState *interp)
 {
     static int busy = 0;
 
-    /* only service pending calls on main thread */
-    if (_PyRuntime.ceval.pending.main_thread &&
-        PyThread_get_thread_ident() != _PyRuntime.ceval.pending.main_thread)
-    {
-        return 0;
-    }
-
     /* don't perform recursive pending calls */
     if (busy) {
         return 0;
@@ -414,13 +451,13 @@ make_pending_calls(void)
     busy = 1;
     /* unsignal before starting to call callbacks, so that any callback
        added in-between re-signals */
-    UNSIGNAL_PENDING_CALLS();
+    UNSIGNAL_PENDING_CALLS(interp);
     int res = 0;
 
-    if (!_PyRuntime.ceval.pending.lock) {
+    if (!interp->ceval.pending.lock) {
         /* initial allocation of the lock */
-        _PyRuntime.ceval.pending.lock = PyThread_allocate_lock();
-        if (_PyRuntime.ceval.pending.lock == NULL) {
+        interp->ceval.pending.lock = PyThread_allocate_lock();
+        if (interp->ceval.pending.lock == NULL) {
             res = -1;
             goto error;
         }
@@ -428,24 +465,18 @@ make_pending_calls(void)
 
     /* perform a bounded number of calls, in case of recursion */
     for (int i=0; i<NPENDINGCALLS; i++) {
-        int j;
-        int (*func)(void *);
+        int (*func)(void *) = NULL;
         void *arg = NULL;
 
         /* pop one item off the queue while holding the lock */
-        PyThread_acquire_lock(_PyRuntime.ceval.pending.lock, WAIT_LOCK);
-        j = _PyRuntime.ceval.pending.first;
-        if (j == _PyRuntime.ceval.pending.last) {
-            func = NULL; /* Queue empty */
-        } else {
-            func = _PyRuntime.ceval.pending.calls[j].func;
-            arg = _PyRuntime.ceval.pending.calls[j].arg;
-            _PyRuntime.ceval.pending.first = (j + 1) % NPENDINGCALLS;
-        }
-        PyThread_release_lock(_PyRuntime.ceval.pending.lock);
+        PyThread_acquire_lock(interp->ceval.pending.lock, WAIT_LOCK);
+        _pop_pending_call(interp, &func, &arg);
+        PyThread_release_lock(interp->ceval.pending.lock);
+
         /* having released the lock, perform the callback */
-        if (func == NULL)
+        if (func == NULL) {
             break;
+        }
         res = func(arg);
         if (res) {
             goto error;
@@ -457,10 +488,18 @@ make_pending_calls(void)
 
 error:
     busy = 0;
-    SIGNAL_PENDING_CALLS();
+    SIGNAL_PENDING_CALLS(interp); /* We're not done yet */
     return res;
 }
 
+int
+_Py_MakePendingCalls(PyInterpreterState *interp)
+{
+    assert(PyGILState_Check());
+
+    return make_pending_calls(interp);
+}
+
 /* Py_MakePendingCalls() is a simple wrapper for the sake
    of backward-compatibility. */
 int
@@ -475,12 +514,8 @@ Py_MakePendingCalls(void)
         return res;
     }
 
-    res = make_pending_calls();
-    if (res != 0) {
-        return res;
-    }
-
-    return 0;
+    PyInterpreterState *interp = _PyRuntime.interpreters.main;
+    return make_pending_calls(interp);
 }
 
 /* The interpreter's recursion limit */
@@ -687,7 +722,7 @@ _PyEval_EvalFrameDefault(PyFrameObject *f, int throwflag)
 
 #define DISPATCH() \
     { \
-        if (!_Py_atomic_load_relaxed(&_PyRuntime.ceval.eval_breaker)) { \
+        if (!_Py_atomic_load_relaxed(&tstate->interp->ceval.eval_breaker)) { \
                     FAST_DISPATCH(); \
         } \
         continue; \
@@ -989,7 +1024,7 @@ main_loop:
            async I/O handler); see Py_AddPendingCall() and
            Py_MakePendingCalls() above. */
 
-        if (_Py_atomic_load_relaxed(&_PyRuntime.ceval.eval_breaker)) {
+        if (_Py_atomic_load_relaxed(&(tstate->interp->ceval.eval_breaker))) {
             opcode = _Py_OPCODE(*next_instr);
             if (opcode == SETUP_FINALLY ||
                 opcode == SETUP_WITH ||
@@ -1022,9 +1057,9 @@ main_loop:
                 }
             }
             if (_Py_atomic_load_relaxed(
-                        &_PyRuntime.ceval.pending.calls_to_do))
+                        &(tstate->interp->ceval.pending.calls_to_do)))
             {
-                if (make_pending_calls() != 0) {
+                if (_Py_MakePendingCalls(tstate->interp) != 0) {
                     goto error;
                 }
             }
@@ -1056,7 +1091,7 @@ main_loop:
             if (tstate->async_exc != NULL) {
                 PyObject *exc = tstate->async_exc;
                 tstate->async_exc = NULL;
-                UNSIGNAL_ASYNC_EXC();
+                UNSIGNAL_ASYNC_EXC(tstate->interp);
                 PyErr_SetNone(exc);
                 Py_DECREF(exc);
                 goto error;
diff --git a/Python/ceval_gil.h b/Python/ceval_gil.h
index f2d5fdb..d9ad361 100644
--- a/Python/ceval_gil.h
+++ b/Python/ceval_gil.h
@@ -176,7 +176,7 @@ static void drop_gil(PyThreadState *tstate)
                     &_PyRuntime.ceval.gil.last_holder)
             ) == tstate)
         {
-        RESET_GIL_DROP_REQUEST();
+        RESET_GIL_DROP_REQUEST(tstate->interp);
             /* NOTE: if COND_WAIT does not atomically start waiting when
                releasing the mutex, another thread can run through, take
                the GIL and drop it again, and reset the condition
@@ -213,7 +213,7 @@ static void take_gil(PyThreadState *tstate)
         if (timed_out &&
             _Py_atomic_load_relaxed(&_PyRuntime.ceval.gil.locked) &&
             _PyRuntime.ceval.gil.switch_number == saved_switchnum) {
-            SET_GIL_DROP_REQUEST();
+            SET_GIL_DROP_REQUEST(tstate->interp);
         }
     }
 _ready:
@@ -239,10 +239,10 @@ _ready:
     MUTEX_UNLOCK(_PyRuntime.ceval.gil.switch_mutex);
 #endif
     if (_Py_atomic_load_relaxed(&_PyRuntime.ceval.gil_drop_request)) {
-        RESET_GIL_DROP_REQUEST();
+        RESET_GIL_DROP_REQUEST(tstate->interp);
     }
     if (tstate->async_exc != NULL) {
-        _PyEval_SignalAsyncExc();
+        _PyEval_SignalAsyncExc(tstate->interp);
     }
 
     MUTEX_UNLOCK(_PyRuntime.ceval.gil.mutex);
diff --git a/Python/pylifecycle.c b/Python/pylifecycle.c
index 8d0075a..eb0fddd 100644
--- a/Python/pylifecycle.c
+++ b/Python/pylifecycle.c
@@ -1459,8 +1459,32 @@ Py_EndInterpreter(PyThreadState *tstate)
     if (tstate->frame != NULL)
         Py_FatalError("Py_EndInterpreter: thread still has a frame");
 
+    // Mark as finalizing.
+    if (interp->ceval.pending.lock != NULL) {
+        PyThread_acquire_lock(interp->ceval.pending.lock, 1);
+    }
+    interp->finalizing = 1;
+    if (interp->ceval.pending.lock != NULL) {
+        PyThread_release_lock(interp->ceval.pending.lock);
+    }
+
+    // Wrap up existing threads.
     wait_for_thread_shutdown();
 
+    // Make any pending calls.
+    if (_Py_atomic_load_relaxed(
+                &(interp->ceval.pending.calls_to_do)))
+    {
+        // XXX Ensure that the interpreter is running in the current thread?
+        if (_Py_MakePendingCalls(interp) < 0) {
+            PyObject *exc, *val, *tb;
+            PyErr_Fetch(&exc, &val, &tb);
+            PyErr_BadInternalCall();
+            _PyErr_ChainExceptions(exc, val, tb);
+            PyErr_Print();
+        }
+    }
+
     call_py_exitfuncs(interp);
 
     if (tstate != interp->tstate_head || tstate->next != NULL)
diff --git a/Python/pystate.c b/Python/pystate.c
index cdb8813..6fe947d 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -132,28 +132,19 @@ PyInterpreterState_New(void)
         return NULL;
     }
 
+    memset(interp, 0, sizeof(*interp));
     interp->id_refcount = -1;
-    interp->id_mutex = NULL;
-    interp->modules = NULL;
-    interp->modules_by_index = NULL;
-    interp->sysdict = NULL;
-    interp->builtins = NULL;
-    interp->builtins_copy = NULL;
-    interp->tstate_head = NULL;
     interp->check_interval = 100;
-    interp->num_threads = 0;
-    interp->pythread_stacksize = 0;
-    interp->codec_search_path = NULL;
-    interp->codec_search_cache = NULL;
-    interp->codec_error_registry = NULL;
-    interp->codecs_initialized = 0;
-    interp->fscodec_initialized = 0;
+
+    interp->ceval.pending.lock = PyThread_allocate_lock();
+    if (interp->ceval.pending.lock == NULL) {
+        PyErr_SetString(PyExc_RuntimeError,
+                        "failed to create interpreter ceval pending mutex");
+        return NULL;
+    }
     interp->core_config = _PyCoreConfig_INIT;
     interp->config = _PyMainInterpreterConfig_INIT;
-    interp->importlib = NULL;
-    interp->import_func = NULL;
     interp->eval_frame = _PyEval_EvalFrameDefault;
-    interp->co_extra_user_count = 0;
 #ifdef HAVE_DLOPEN
 #if HAVE_DECL_RTLD_NOW
     interp->dlopenflags = RTLD_NOW;
@@ -161,13 +152,10 @@ PyInterpreterState_New(void)
     interp->dlopenflags = RTLD_LAZY;
 #endif
 #endif
-#ifdef HAVE_FORK
-    interp->before_forkers = NULL;
-    interp->after_forkers_parent = NULL;
-    interp->after_forkers_child = NULL;
-#endif
-    interp->pyexitfunc = NULL;
-    interp->pyexitmodule = NULL;
+
+    if (_PyRuntime.main_thread == 0) {
+        _PyRuntime.main_thread = PyThread_get_thread_ident();
+    }
 
     HEAD_LOCK();
     if (_PyRuntime.interpreters.next_id < 0) {
@@ -222,6 +210,9 @@ PyInterpreterState_Clear(PyInterpreterState *interp)
     Py_CLEAR(interp->after_forkers_parent);
     Py_CLEAR(interp->after_forkers_child);
 #endif
+    // 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.
 }
 
 
@@ -262,6 +253,9 @@ PyInterpreterState_Delete(PyInterpreterState *interp)
     if (interp->id_mutex != NULL) {
         PyThread_free_lock(interp->id_mutex);
     }
+    if (interp->ceval.pending.lock != NULL) {
+        PyThread_free_lock(interp->ceval.pending.lock);
+    }
     PyMem_RawFree(interp);
 }
 
@@ -871,7 +865,7 @@ PyThreadState_SetAsyncExc(unsigned long id, PyObject *exc)
             p->async_exc = exc;
             HEAD_UNLOCK();
             Py_XDECREF(old_exc);
-            _PyEval_SignalAsyncExc();
+            _PyEval_SignalAsyncExc(interp);
             return 1;
         }
     }
@@ -1338,6 +1332,7 @@ _PyCrossInterpreterData_Release(_PyCrossInterpreterData *data)
     }
 
     // "Release" the data and/or the object.
+    // XXX Use _Py_AddPendingCall().
     _call_in_interpreter(interp, _release_xidata, data);
 }