Refactor server deallocation

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

@@ -49,7 +49,7 @@ cdef grpc_event _next(grpc_completion_queue *c_completion_queue, deadline):
 cdef _interpret_event(grpc_event c_event):
   cdef _Tag tag
   if c_event.type == GRPC_QUEUE_TIMEOUT:
-    # NOTE(nathaniel): For now we coopt ConnectivityEvent here.
+    # TODO(ericgribkoff) Do not coopt ConnectivityEvent here.
     return None, ConnectivityEvent(GRPC_QUEUE_TIMEOUT, False, None)
   elif c_event.type == GRPC_QUEUE_SHUTDOWN:
     # NOTE(nathaniel): For now we coopt ConnectivityEvent here.

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

@@ -128,7 +128,9 @@ cdef class Server:
       with nogil:
         grpc_server_cancel_all_calls(self.c_server)
 
-  def __dealloc__(self):
+  # TODO(ericgribkoff) Determine what, if any, portion of this is safe to call
+  # from __dealloc__, and potentially remove backup_shutdown_queue.
+  def destroy(self):
     if self.c_server != NULL:
       if not self.is_started:
         pass
@@ -146,4 +148,8 @@ cdef class Server:
         while not self.is_shutdown:
           time.sleep(0)
       grpc_server_destroy(self.c_server)
-    grpc_shutdown()
+      self.c_server = NULL
+
+  def __dealloc(self):
+    if self.c_server == NULL:
+      grpc_shutdown()

src/python/grpcio/grpc/_server.py

@@ -48,7 +48,7 @@ _CANCELLED = 'cancelled'
 
 _EMPTY_FLAGS = 0
 
-_UNEXPECTED_EXIT_SERVER_GRACE = 1.0
+_DEALLOCATED_SERVER_CHECK_PERIOD_S = 1.0
 
 
 def _serialized_request(request_event):
@@ -676,6 +676,9 @@ class _ServerState(object):
         self.rpc_states = set()
         self.due = set()
 
+        # A "volatile" flag to interrupt the daemon serving thread
+        self.server_deallocated = False
+
 
 def _add_generic_handlers(state, generic_handlers):
     with state.lock:
@@ -702,6 +705,7 @@ def _request_call(state):
 # TODO(https://github.com/grpc/grpc/issues/6597): delete this function.
 def _stop_serving(state):
     if not state.rpc_states and not state.due:
+        state.server.destroy()
         for shutdown_event in state.shutdown_events:
             shutdown_event.set()
         state.stage = _ServerStage.STOPPED
@@ -715,49 +719,69 @@ def _on_call_completed(state):
         state.active_rpc_count -= 1
 
 
-def _serve(state):
+def _process_event_and_continue(state, event):
-    while True:
+    should_continue = True
-        event = state.completion_queue.poll()
+    if event.tag is _SHUTDOWN_TAG:
-        if event.tag is _SHUTDOWN_TAG:
+        with state.lock:
+            state.due.remove(_SHUTDOWN_TAG)
+            if _stop_serving(state):
+                should_continue = False
+    elif event.tag is _REQUEST_CALL_TAG:
+        with state.lock:
+            state.due.remove(_REQUEST_CALL_TAG)
+            concurrency_exceeded = (
+                state.maximum_concurrent_rpcs is not None and
+                state.active_rpc_count >= state.maximum_concurrent_rpcs)
+            rpc_state, rpc_future = _handle_call(
+                event, state.generic_handlers, state.interceptor_pipeline,
+                state.thread_pool, concurrency_exceeded)
+            if rpc_state is not None:
+                state.rpc_states.add(rpc_state)
+            if rpc_future is not None:
+                state.active_rpc_count += 1
+                rpc_future.add_done_callback(
+                    lambda unused_future: _on_call_completed(state))
+            if state.stage is _ServerStage.STARTED:
+                _request_call(state)
+            elif _stop_serving(state):
+                should_continue = False
+    else:
+        rpc_state, callbacks = event.tag(event)
+        for callback in callbacks:
+            callable_util.call_logging_exceptions(callback,
+                                                  'Exception calling callback!')
+        if rpc_state is not None:
             with state.lock:
-                state.due.remove(_SHUTDOWN_TAG)
+                state.rpc_states.remove(rpc_state)
                 if _stop_serving(state):
-                    return
+                    should_continue = False
-        elif event.tag is _REQUEST_CALL_TAG:
+    return should_continue
-            with state.lock:
+
-                state.due.remove(_REQUEST_CALL_TAG)
+
-                concurrency_exceeded = (
+def _serve(state):
-                    state.maximum_concurrent_rpcs is not None and
+    while True:
-                    state.active_rpc_count >= state.maximum_concurrent_rpcs)
+        timeout = time.time() + _DEALLOCATED_SERVER_CHECK_PERIOD_S
-                rpc_state, rpc_future = _handle_call(
+        event = state.completion_queue.poll(timeout)
-                    event, state.generic_handlers, state.interceptor_pipeline,
+        if state.server_deallocated:
-                    state.thread_pool, concurrency_exceeded)
+            _begin_shutdown_once(state)
-                if rpc_state is not None:
+        if event.completion_type != cygrpc.CompletionType.queue_timeout:
-                    state.rpc_states.add(rpc_state)
+            if not _process_event_and_continue(state, event):
-                if rpc_future is not None:
+                return
-                    state.active_rpc_count += 1
-                    rpc_future.add_done_callback(
-                        lambda unused_future: _on_call_completed(state))
-                if state.stage is _ServerStage.STARTED:
-                    _request_call(state)
-                elif _stop_serving(state):
-                    return
-        else:
-            rpc_state, callbacks = event.tag(event)
-            for callback in callbacks:
-                callable_util.call_logging_exceptions(
-                    callback, 'Exception calling callback!')
-            if rpc_state is not None:
-                with state.lock:
-                    state.rpc_states.remove(rpc_state)
-                    if _stop_serving(state):
-                        return
         # We want to force the deletion of the previous event
         # ~before~ we poll again; if the event has a reference
         # to a shutdown Call object, this can induce spinlock.
         event = None
 
 
+def _begin_shutdown_once(state):
+    with state.lock:
+        if state.stage is _ServerStage.STARTED:
+            state.server.shutdown(state.completion_queue, _SHUTDOWN_TAG)
+            state.stage = _ServerStage.GRACE
+            state.shutdown_events = []
+            state.due.add(_SHUTDOWN_TAG)
+
+
 def _stop(state, grace):
     with state.lock:
         if state.stage is _ServerStage.STOPPED:
@@ -765,11 +789,7 @@ def _stop(state, grace):
             shutdown_event.set()
             return shutdown_event
         else:
-            if state.stage is _ServerStage.STARTED:
+            _begin_shutdown_once(state)
-                state.server.shutdown(state.completion_queue, _SHUTDOWN_TAG)
-                state.stage = _ServerStage.GRACE
-                state.shutdown_events = []
-                state.due.add(_SHUTDOWN_TAG)
             shutdown_event = threading.Event()
             state.shutdown_events.append(shutdown_event)
             if grace is None:
@@ -840,7 +860,9 @@ class _Server(grpc.Server):
         return _stop(self._state, grace)
 
     def __del__(self):
-        _stop(self._state, None)
+        # We can not grab a lock in __del__(), so set a flag to signal the
+        # serving daemon thread (if it exists) to initiate shutdown.
+        self._state.server_deallocated = True
 
 
 def create_server(thread_pool, generic_rpc_handlers, interceptors, options,

src/python/grpcio_tests/tests/channelz/_channelz_servicer_test.py

@@ -88,8 +88,6 @@ def _generate_channel_server_pairs(n):
 def _close_channel_server_pairs(pairs):
     for pair in pairs:
         pair.server.stop(None)
-        # TODO(ericgribkoff) This del should not be required
-        del pair.server
         pair.channel.close()