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.

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;
@@ -24,6 +28,7 @@ struct _pending_calls {
     int async_exc;
 #define NPENDINGCALLS 32
     struct {
+        unsigned long thread_id;
         int (*func)(void *);
         void *arg;
     } calls[NPENDINGCALLS];
@@ -31,6 +36,13 @@ 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 {
@@ -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 @@ class TestPendingCalls(unittest.TestCase):
     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

+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

@@ -1146,7 +1146,7 @@ Py_FinalizeEx(void)
     interp = tstate->interp;
 
     // Make any remaining pending calls.
-    _Py_FinishPendingCalls();
+    _Py_FinishPendingCalls(interp);
 
     /* The interpreter is still entirely intact at this point, and the
      * exit funcs may be relying on that.  In particular, if some thread
@@ -1552,6 +1552,9 @@ Py_EndInterpreter(PyThreadState *tstate)
     // Wrap up existing "threading"-module-created, non-daemon threads.
     wait_for_thread_shutdown();
 
+    // Make any remaining pending calls.
+    _Py_FinishPendingCalls(interp);
+
     call_py_exitfuncs(interp);
 
     if (tstate != interp->tstate_head || tstate->next != NULL)

Python/pystate.c

@@ -173,6 +173,14 @@ PyInterpreterState_New(void)
     memset(interp, 0, sizeof(*interp));
     interp->id_refcount = -1;
     interp->check_interval = 100;
+
+    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->eval_frame = _PyEval_EvalFrameDefault;
 #ifdef HAVE_DLOPEN
@@ -279,6 +287,9 @@ 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);
 }
 
@@ -928,7 +939,7 @@ PyThreadState_SetAsyncExc(unsigned long id, PyObject *exc)
             p->async_exc = exc;
             HEAD_UNLOCK();
             Py_XDECREF(old_exc);
-            _PyEval_SignalAsyncExc();
+            _PyEval_SignalAsyncExc(interp);
             return 1;
         }
     }
@@ -1342,7 +1353,7 @@ _PyObject_GetCrossInterpreterData(PyObject *obj, _PyCrossInterpreterData *data)
     return 0;
 }
 
-static void
+static int
 _release_xidata(void *arg)
 {
     _PyCrossInterpreterData *data = (_PyCrossInterpreterData *)arg;
@@ -1350,30 +1361,8 @@ _release_xidata(void *arg)
         data->free(data->data);
     }
     Py_XDECREF(data->obj);
-}
-
-static void
-_call_in_interpreter(PyInterpreterState *interp,
-                     void (*func)(void *), void *arg)
-{
-    /* We would use Py_AddPendingCall() if it weren't specific to the
-     * main interpreter (see bpo-33608).  In the meantime we take a
-     * naive approach.
-     */
-    PyThreadState *save_tstate = NULL;
-    if (interp != _PyInterpreterState_Get()) {
-        // XXX Using the "head" thread isn't strictly correct.
-        PyThreadState *tstate = PyInterpreterState_ThreadHead(interp);
-        // XXX Possible GILState issues?
-        save_tstate = PyThreadState_Swap(tstate);
-    }
-
-    func(arg);
-
-    // Switch back.
-    if (save_tstate != NULL) {
-        PyThreadState_Swap(save_tstate);
-    }
+    PyMem_Free(data);
+    return 0;
 }
 
 void
@@ -1384,7 +1373,7 @@ _PyCrossInterpreterData_Release(_PyCrossInterpreterData *data)
         return;
     }
 
-    // Switch to the original interpreter.
+    // Get the original interpreter.
     PyInterpreterState *interp = _PyInterpreterState_LookUpID(data->interp);
     if (interp == NULL) {
         // The intepreter was already destroyed.
@@ -1393,9 +1382,24 @@ _PyCrossInterpreterData_Release(_PyCrossInterpreterData *data)
         }
         return;
     }
+    // XXX There's an ever-so-slight race here...
+    if (interp->finalizing) {
+        // XXX Someone leaked some memory...
+        return;
+    }
 
     // "Release" the data and/or the object.
-    _call_in_interpreter(interp, _release_xidata, data);
+    _PyCrossInterpreterData *copied = PyMem_Malloc(sizeof(_PyCrossInterpreterData));
+    if (copied == NULL) {
+        PyErr_SetString(PyExc_MemoryError,
+                        "Not enough memory to preserve cross-interpreter data");
+        PyErr_Print();
+        return;
+    }
+    memcpy(copied, data, sizeof(_PyCrossInterpreterData));
+    if (_Py_AddPendingCall(interp, 0, _release_xidata, copied) != 0) {
+        // XXX Queue full or couldn't get lock.  Try again somehow?
+    }
 }
 
 PyObject *