Fix when the Timekeeper is destroyed with in-flight delays

Summary:
[Folly] Fix when the Timekeeper is destroyed with in-flight delays.

Previously, the Timekeeper dtor would immediately trigger the delayed callbacks early with a success result. But there are various situations where that is wrong behavior, such as when the callback enqueues more delayed callbacks in the success path. Higher-level, a future returned by `Timekeeper::after` or `sleep` would be completed early with a value.

Now, let it immediately trigger the delayed callbacks early with an error result. Higher-level, a future returned by `Timekeeper::after` or `sleep` will be completed early with an error.

Also: use `NoTimeout` consistently, and, on interrupt, pass the interrupt exception through.

Differential Revision: D6837669

fbshipit-source-id: 7d18d7f57999d56bbb5269cacdb443ac22448cf0

folly/futures/ThreadWheelTimekeeper.cpp

@@ -67,6 +67,14 @@ struct WTCallback : public std::enable_shared_from_this<WTCallback>,
     }
   }
 
+  void callbackCanceled() noexcept override {
+    // Don't need Promise anymore, break the circular reference
+    auto promise = stealPromise();
+    if (!promise.isFulfilled()) {
+      promise.setException(NoTimekeeper{});
+    }
+  }
+
   void interruptHandler(exception_wrapper ew) {
     // Capture shared_ptr of self in lambda, if we don't do this, object
     // may go away before the lambda is executed from event base thread.

folly/futures/test/TimekeeperTest.cpp

@@ -253,3 +253,13 @@ TEST_F(TimekeeperFixture, howToCastDuration) {
   auto f = timeLord_->after(std::chrono::duration_cast<Duration>(
       std::chrono::nanoseconds(1)));
 }
+
+TEST_F(TimekeeperFixture, destruction) {
+  folly::Optional<ThreadWheelTimekeeper> tk;
+  tk.emplace();
+  auto f = tk->after(std::chrono::seconds(10));
+  EXPECT_FALSE(f.isReady());
+  tk.clear();
+  EXPECT_TRUE(f.isReady());
+  EXPECT_TRUE(f.hasException());
+}