tcp/context: use mutex to avoid race condition when multiple operations are running

Differential Revision: D72599396

Pull Request resolved: #427

Author: d4l3k

gloo/transport/tcp/context.cc

@@ -177,6 +177,8 @@ std::unique_ptr<transport::Pair>& Context::getPair(int rank) {
     return pair;
   }
 
+  std::lock_guard<std::mutex> lock(m_);
+
   if (!connecting_[rank]) {
     connecting_[rank] = true;
 
@@ -187,6 +189,7 @@ std::unique_ptr<transport::Pair>& Context::getPair(int rank) {
 
     auto remoteDeviceAddr = Address(remoteRankInfo.addressBytes).getSockaddr();
     auto remoteAddr = Address(remoteDeviceAddr, this->rank);
+    // Actual connection happens asynchronously.
     pair->connect(remoteAddr.bytes());
   }
   return pair;

gloo/transport/tcp/context.h

@@ -52,6 +52,10 @@ class Context : public ::gloo::transport::Context,
  protected:
   std::shared_ptr<Device> device_;
   std::shared_ptr<IStore> store_{nullptr};
+
+  // Protects the connection states to avoid race conditions.
+  std::mutex m_;
+  // Whether or not connection has been started for this peer.
   std::vector<bool> connecting_;
 
   using pendingRecvTuple = std::tuple<