Fix assumption in mojo binding class Connector that message pipes notice the other end's closing synchronously.

This was causing mojo_public_bindings_unittests' MultiplexRouterTest.BasicRequestResponse and MultiplexRouterTest.RequestWithNoReceiver to fail with the new EDK. The problem is that in Connector::HandleError message_pipe_ is reset to a dummy message pipe whose other end is closed immediately. In the old EDK, the live end notices that the peer is closed synchronously. In the new EDK this happens asynchronously because of thread hops to the IO thread. The fix is to simply ensure we don't watch the message pipe handle twice (since it's not cancelled immediately).

BUG=561803

Review URL: https://codereview.chromium.org/1557753002

Cr-Commit-Position: refs/heads/master@{#367327}

1 files changed, 7 insertions, 1 deletions

diff --git a/mojo/public/cpp/bindings/lib/connector.cc b/mojo/public/cpp/bindings/lib/connector.cc
index c5e9b7f..e6f2d83 100644
--- a/mojo/public/cpp/bindings/lib/connector.cc
+++ b/mojo/public/cpp/bindings/lib/connector.cc
@@ -264,7 +264,13 @@ void Connector::ReadAllAvailableMessages() {
       return;
 
     if (rv == MOJO_RESULT_SHOULD_WAIT) {
-      WaitToReadMore();
+      // ReadSingleMessage could end up calling HandleError which resets
+      // message_pipe_ to a dummy one that is closed. The old EDK will see the
+      // that the peer is closed immediately, while the new one is asynchronous
+      // because of thread hops. In that case, there'll still be an async
+      // waiter.
+      if (!async_wait_id_)
+        WaitToReadMore();
       break;
     }
   }