From f4a9b9acbaa1dcb2bb9aa46b6aede6da4d70dfc4 Mon Sep 17 00:00:00 2001
From: gilbertlee-amd <44450918+gilbertlee-amd@users.noreply.github.com>
Date: Fri, 26 Feb 2021 16:29:55 -0700
Subject: [PATCH] Adding pthread_join / pthread_detach to clean up pthreads to
 avoid leaks (#322)

---
 src/bootstrap.cc        | 5 ++++-
 src/transport/net_ib.cc | 4 +++-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/bootstrap.cc b/src/bootstrap.cc
index b4db1368fd..d7a50e26a0 100644
--- a/src/bootstrap.cc
+++ b/src/bootstrap.cc
@@ -194,6 +194,7 @@ ncclResult_t bootstrapCreateRoot(ncclUniqueId* id, bool idFromEnv) {
   rootStruct->hash = djb2Hash(id->internal);
   rootStruct->listenFd = listenFd;
   pthread_create(&thread, NULL, bootstrapRoot, (void *)rootStruct);
+  pthread_detach(thread); // [RCCL] Adding detach to properly clean up bootstrapRoot thread
   // [/RCCL]
 
   return ncclSuccess;
@@ -533,7 +534,9 @@ ncclResult_t bootstrapClose(void* commState) {
   state->allocState->stop = 1;
 
   // Join the allocThread so we catch resource leaks as being hung here
-  // pthread_join(state->allocThread, nullptr);
+  // [RCCL] Uncommenting this join to clean up the allocThread
+  pthread_join(state->allocThread, nullptr);
+  // [/RCCL]
 
   free(state->peerCommAddresses);
   free(state->peerAllocAddresses);
diff --git a/src/transport/net_ib.cc b/src/transport/net_ib.cc
index 228a3fd4f8..73893cfb72 100644
--- a/src/transport/net_ib.cc
+++ b/src/transport/net_ib.cc
@@ -186,6 +186,9 @@ ncclResult_t ncclIbInit(ncclDebugLogger_t logFunction) {
           ncclNIbDevs++;
           nPorts++;
           pthread_create(&ncclIbAsyncThread, NULL, ncclIbAsyncThreadMain, context);
+          // [RCCL]
+          pthread_detach(ncclIbAsyncThread);
+          // [/RCCL]
         }
         if (nPorts == 0 && ncclSuccess != wrap_ibv_close_device(context)) { return ncclInternalError; }
       }
@@ -915,4 +918,3 @@ ncclNet_t ncclNetIb = {
   ncclIbCloseRecv,
   ncclIbCloseListen
 };
-