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

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. */

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;

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;

Misc/NEWS.d/next/Core and Builtins/2018-09-15-12-13-46.bpo-33608.avmvVP.rst

+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.

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
 

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);
                 }
             }
         }

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);

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)

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);
 }