fix: fix racy conditions during shutdown (googleapis/google-cloud-cpp…

…-common#250)

* bug: fix race condition during shutdown

We were checking if the completion queue was shutdown and then
scheduling work, but not atomically.

google/cloud/completion_queue.cc

@@ -90,10 +90,8 @@ google::cloud::future<StatusOr<std::chrono::system_clock::time_point>>
 CompletionQueue::MakeDeadlineTimer(
     std::chrono::system_clock::time_point deadline) {
   auto op = std::make_shared<AsyncTimerFuture>(impl_->CreateAlarm());
-  void* tag = impl_->RegisterOperation(op);
-  if (tag != nullptr) {
-    op->Set(impl_->cq(), deadline, tag);
-  }
+  impl_->StartOperation(
+      op, [&](void* tag) { op->Set(impl_->cq(), deadline, tag); });
   return op->GetFuture();
 }
 

google/cloud/completion_queue.h

@@ -113,10 +113,9 @@ class CompletionQueue {
       std::unique_ptr<grpc::ClientContext> context) {
     auto op =
         std::make_shared<internal::AsyncUnaryRpcFuture<Request, Response>>();
-    void* tag = impl_->RegisterOperation(op);
-    if (tag != nullptr) {
+    impl_->StartOperation(op, [&](void* tag) {
       op->Start(async_call, std::move(context), request, &impl_->cq(), tag);
-    }
+    });
     return op->GetFuture();
   }
 

google/cloud/completion_queue_test.cc

@@ -333,11 +333,14 @@ TEST(CompletionQueueTest, RunAsync) {
 // Sets up a timer that reschedules itself and verifies we can shut down
 // cleanly whether we call `CancelAll()` on the queue first or not.
 namespace {
-void RunAndReschedule(CompletionQueue& cq, bool ok) {
+using TimerFuture = future<StatusOr<std::chrono::system_clock::time_point>>;
+
+void RunAndReschedule(CompletionQueue& cq, bool ok,
+                      std::chrono::seconds duration) {
   if (ok) {
-    cq.MakeRelativeTimer(std::chrono::seconds(1))
-        .then([&cq](future<StatusOr<std::chrono::system_clock::time_point>>
-                        result) { RunAndReschedule(cq, result.get().ok()); });
+    cq.MakeRelativeTimer(duration).then([&cq, duration](TimerFuture result) {
+      RunAndReschedule(cq, result.get().ok(), duration);
+    });
   }
 }
 }  // namespace
@@ -346,17 +349,40 @@ TEST(CompletionQueueTest, ShutdownWithReschedulingTimer) {
   CompletionQueue cq;
   std::thread t([&cq] { cq.Run(); });
 
-  RunAndReschedule(cq, /*ok=*/true);
+  RunAndReschedule(cq, /*ok=*/true, std::chrono::seconds(1));
 
   cq.Shutdown();
   t.join();
 }
 
+TEST(CompletionQueueTest, ShutdownWithFastReschedulingTimer) {
+  auto constexpr kThreadCount = 32;
+  auto constexpr kTimerCount = 100;
+  CompletionQueue cq;
+  std::vector<std::thread> threads(kThreadCount);
+  std::generate_n(threads.begin(), threads.size(),
+                  [&cq] { return std::thread([&cq] { cq.Run(); }); });
+
+  for (int i = 0; i != kTimerCount; ++i) {
+    RunAndReschedule(cq, /*ok=*/true, std::chrono::seconds(0));
+  }
+
+  promise<void> wait;
+  cq.MakeRelativeTimer(std::chrono::milliseconds(1)).then([&wait](TimerFuture) {
+    wait.set_value();
+  });
+  wait.get_future().get();
+  cq.Shutdown();
+  for (auto& t : threads) {
+    t.join();
+  }
+}
+
 TEST(CompletionQueueTest, CancelAndShutdownWithReschedulingTimer) {
   CompletionQueue cq;
   std::thread t([&cq] { cq.Run(); });
 
-  RunAndReschedule(cq, /*ok=*/true);
+  RunAndReschedule(cq, /*ok=*/true, std::chrono::seconds(1));
 
   cq.CancelAll();
   cq.Shutdown();

google/cloud/internal/async_read_stream_impl.h

@@ -157,15 +157,14 @@ class AsyncReadStreamImpl
     context_ = std::move(context);
     cq_ = std::move(cq);
     auto callback = std::make_shared<NotifyStart>(this->shared_from_this());
-    void* tag = cq_->RegisterOperation(std::move(callback));
-    // @note If `tag == nullptr` the `CompletionQueue` has been `Shutdown()`.
-    //     We leave `reader_` null in this case; other methods must make the
-    //     same `tag != nullptr` check prior to accessing `reader_`.  This is
-    //     safe since `Shutdown()` cannot be undone.
-    if (tag != nullptr) {
+    cq_->StartOperation(std::move(callback), [&](void* tag) {
+      // @note If the the `CompletionQueue` has been `Shutdown()` this lambda is
+      //     never called. We leave `reader_` null in this case; other methods
+      //     must make the same `tag != nullptr` check prior to accessing
+      //     `reader_`.  This is safe since `Shutdown()` cannot be undone.
       reader_ = async_call(context_.get(), request, &cq_->cq());
       reader_->StartCall(tag);
-    }
+    });
   }
 
   /// Cancel the current streaming read RPC.
@@ -202,10 +201,8 @@ class AsyncReadStreamImpl
 
     auto callback = std::make_shared<NotifyRead>(this->shared_from_this());
     auto response = &callback->response;
-    void* tag = cq_->RegisterOperation(std::move(callback));
-    if (tag != nullptr) {
-      reader_->Read(response, tag);
-    }
+    cq_->StartOperation(std::move(callback),
+                        [&](void* tag) { reader_->Read(response, tag); });
   }
 
   /// Handle the result of a `Read()` call.
@@ -252,10 +249,8 @@ class AsyncReadStreamImpl
 
     auto callback = std::make_shared<NotifyFinish>(this->shared_from_this());
     auto status = &callback->status;
-    void* tag = cq_->RegisterOperation(std::move(callback));
-    if (tag != nullptr) {
-      reader_->Finish(status, tag);
-    }
+    cq_->StartOperation(std::move(callback),
+                        [&](void* tag) { reader_->Finish(status, tag); });
   }
 
   /// Handle the result of a Finish() request.
@@ -292,10 +287,8 @@ class AsyncReadStreamImpl
 
     auto callback = std::make_shared<NotifyDiscard>(this->shared_from_this());
     auto response = &callback->response;
-    void* tag = cq_->RegisterOperation(std::move(callback));
-    if (tag != nullptr) {
-      reader_->Read(response, tag);
-    }
+    cq_->StartOperation(std::move(callback),
+                        [&](void* tag) { reader_->Read(response, tag); });
   }
 
   /// Handle the result of a Discard() call.

google/cloud/internal/completion_queue_impl.cc

@@ -73,26 +73,6 @@ std::unique_ptr<grpc::Alarm> CompletionQueueImpl::CreateAlarm() const {
   return google::cloud::internal::make_unique<grpc::Alarm>();
 }
 
-void* CompletionQueueImpl::RegisterOperation(
-    std::shared_ptr<AsyncGrpcOperation> op) {
-  void* tag = op.get();
-  std::unique_lock<std::mutex> lk(mu_);
-  if (shutdown_) {
-    lk.unlock();
-    op->Notify(/*ok=*/false);
-    return nullptr;
-  }
-  auto ins =
-      pending_ops_.emplace(reinterpret_cast<std::intptr_t>(tag), std::move(op));
-  // After this point we no longer need the lock, so release it.
-  lk.unlock();
-  if (ins.second) {
-    return tag;
-  }
-  google::cloud::internal::ThrowRuntimeError(
-      "assertion failure: insertion should succeed");
-}
-
 std::shared_ptr<AsyncGrpcOperation> CompletionQueueImpl::FindOperation(
     void* tag) {
   std::lock_guard<std::mutex> lk(mu_);

google/cloud/internal/completion_queue_impl.h

@@ -247,8 +247,30 @@ class CompletionQueueImpl {
   /// The underlying gRPC completion queue.
   grpc::CompletionQueue& cq() { return cq_; }
 
-  /// Add a new asynchronous operation to the completion queue.
-  void* RegisterOperation(std::shared_ptr<AsyncGrpcOperation> op);
+  /// Atomically add a new operation to the completion queue and start it.
+  template <typename Callable,
+            typename std::enable_if<
+                google::cloud::internal::is_invocable<Callable, void*>::value,
+                int>::type = 0>
+  void StartOperation(std::shared_ptr<AsyncGrpcOperation> op,
+                      Callable&& start) {
+    void* tag = op.get();
+    std::unique_lock<std::mutex> lk(mu_);
+    if (shutdown_) {
+      lk.unlock();
+      op->Notify(/*ok=*/false);
+      return;
+    }
+    auto ins = pending_ops_.emplace(reinterpret_cast<std::intptr_t>(tag),
+                                    std::move(op));
+    if (ins.second) {
+      start(tag);
+      lk.unlock();
+      return;
+    }
+    google::cloud::internal::ThrowRuntimeError(
+        "assertion failure: insertion should succeed");
+  }
 
  protected:
   /// Return the asynchronous operation associated with @p tag.