bpo-33608: Factor out a private, per-interpreter _Py_AddPendingCall(). (gh-12360)

This is effectively an un-revert of #11617 and #12024 (reverted in #12159). Portions of those were merged in other PRs (with lower risk) and this represents the remainder. Note that I found 3 different bugs in the original PRs and have fixed them here.

Author: ericsnowcurrently

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

@@ -11,7 +11,11 @@ extern "C" {
 #include "pycore_atomic.h"
 #include "pythread.h"
 
-PyAPI_FUNC(void) _Py_FinishPendingCalls(void);
+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*);
+PyAPI_FUNC(void) _Py_FinishPendingCalls(struct _is*);
 
 struct _pending_calls {
     int finishing;
@@ -24,13 +28,21 @@ struct _pending_calls {
     int async_exc;
 #define NPENDINGCALLS 32
     struct {
+        unsigned long thread_id;
         int (*func)(void *);
         void *arg;
     } calls[NPENDINGCALLS];
     int first;
     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 {
@@ -41,12 +53,8 @@ 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

@@ -12,6 +12,7 @@ extern "C" {
 #include "pystate.h"
 #include "pythread.h"
 
+#include "pycore_atomic.h"
 #include "pycore_ceval.h"
 #include "pycore_pathconfig.h"
 #include "pycore_pymem.h"
@@ -83,6 +84,8 @@ struct _is {
     PyObject *pyexitmodule;
 
     uint64_t tstate_next_unique_id;
+
+    struct _ceval_interpreter_state ceval;
 };
 
 PyAPI_FUNC(struct _is*) _PyInterpreterState_LookUpID(PY_INT64_T);

Lib/test/test_capi.py

@@ -373,7 +373,7 @@ def pendingcalls_wait(self, l, n, context = None):
     def test_pendingcalls_threaded(self):
 
         #do every callback on a separate thread
-        n = 32 #total callbacks
+        n = 32 #total callbacks (see NPENDINGCALLS in pycore_ceval.h)
         threads = []
         class foo(object):pass
         context = foo()

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.

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);
                 }
             }
         }