Fix a deadlock

This deadlock was introduced when updating sendrecv_test.cu for Connection being returned in a future.
microsoft · Aug 31, 2023 · 87a034a · 87a034a
1 parent 60d54b9
commit 87a034a
Showing 1 changed file with 4 additions and 4 deletions.
diff --git a/test/mscclpp-test/sendrecv_test.cu b/test/mscclpp-test/sendrecv_test.cu
@@ -157,16 +157,16 @@ void SendRecvTestEngine::setupConnections() {
 
   auto sendConnFuture =
       comm_->connectOnSetup(sendToRank, 0, getTransport(args_.rank, sendToRank, args_.nRanksPerNode, ibDevice));
-  comm_->setup();
-  smSemaphores.push_back(std::make_shared<mscclpp::SmDevice2DeviceSemaphore>(*comm_, sendConnFuture.get()));
   if (recvFromRank != sendToRank) {
     auto recvConnFuture =
         comm_->connectOnSetup(recvFromRank, 0, getTransport(args_.rank, recvFromRank, args_.nRanksPerNode, ibDevice));
     comm_->setup();
+    smSemaphores.push_back(std::make_shared<mscclpp::SmDevice2DeviceSemaphore>(*comm_, sendConnFuture.get()));
     smSemaphores.push_back(std::make_shared<mscclpp::SmDevice2DeviceSemaphore>(*comm_, recvConnFuture.get()));
   } else {
-    // reuse the send channel if worldSize is 2
-    smSemaphores.push_back(smSemaphores[0]);
+    comm_->setup();
+    smSemaphores.push_back(std::make_shared<mscclpp::SmDevice2DeviceSemaphore>(*comm_, sendConnFuture.get()));
+    smSemaphores.push_back(smSemaphores[0]);  // reuse the send channel if worldSize is 2
   }
   comm_->setup();