Moved grpc_shutdown to end of Py_Finalize()

We currently rely on the __del__ method of a module scope object
to call grpc_shutdown().  __del__ methods are not guaranteed
to be called, and furthermore there are no guarantees about
ordering, leading to shutdown race conditions.  This moves
grpc_shutdown to Py_Finalize(), which gets called after the
Python context is completely cleaned up.

src/python/grpcio/grpc/_cython/_cygrpc/grpc.pxi

@@ -37,6 +37,7 @@ cdef extern from "grpc/_cython/loader.h":
   ctypedef long int64_t
 
   int pygrpc_load_core(char*)
+  int pygrpc_initialize_core()
 
   void *gpr_malloc(size_t size) nogil
   void gpr_free(void *ptr) nogil

src/python/grpcio/grpc/_cython/cygrpc.pyx

@@ -45,30 +45,20 @@ include "grpc/_cython/_cygrpc/security.pyx.pxi"
 include "grpc/_cython/_cygrpc/server.pyx.pxi"
 
 #
-# Global state
+# initialize gRPC
 #
 
-cdef class _ModuleState:
 
-  cdef bint is_loaded
+def _initialize():
+  if 'win32' in sys.platform:
+    filename = pkg_resources.resource_filename(
+        'grpc._cython', '_windows/grpc_c.64.python')
+    if not pygrpc_load_core(filename):
+      raise ImportError('failed to load core gRPC library')
+  if not pygrpc_initialize_core():
+    raise ImportError('failed to initialize core gRPC library')
 
-  def __cinit__(self):
-    if 'win32' in sys.platform:
-      filename = pkg_resources.resource_filename(
-          'grpc._cython', '_windows/grpc_c.64.python')
-      if not pygrpc_load_core(filename):
-        raise ImportError('failed to load core gRPC library')
-    with nogil:
-      grpc_init()
-    self.is_loaded = True
-    with nogil:
-      grpc_set_ssl_roots_override_callback(
+  grpc_set_ssl_roots_override_callback(
           <grpc_ssl_roots_override_callback>ssl_roots_override_callback)
 
-  def __dealloc__(self):
-    if self.is_loaded:
-      with nogil:
-        grpc_shutdown()
-
-_module_state = _ModuleState()
-
+_initialize()

src/python/grpcio/grpc/_cython/loader.c

@@ -31,6 +31,7 @@
  *
  */
 
+#include <Python.h>
 #include "loader.h"
 
 #ifdef __cplusplus
@@ -62,6 +63,12 @@ int pygrpc_load_core(char *path) { return 1; }
 
 #endif  /* !GPR_WINDOWS */
 
+// Cython doesn't have Py_AtExit bindings, so we call the C_API directly
+int pygrpc_initialize_core(void) {
+  grpc_init();
+  return Py_AtExit(grpc_shutdown) < 0 ? 0 : 1;
+}
+
 #ifdef __cplusplus
 }
 #endif  /* __cpluslus */

src/python/grpcio/grpc/_cython/loader.h

@@ -46,6 +46,11 @@ extern "C" {
 /* Attempts to load the core if necessary, and return non-zero upon succes. */
 int pygrpc_load_core(char *path);
 
+/* Initializes grpc and registers grpc_shutdown() to be called right before
+ * interpreter exit.  Returns non-zero upon success.
+ */
+int pygrpc_initialize_core(void);
+
 #ifdef __cplusplus
 }
 #endif  /* __cpluslus */