From b7b7011def413b2be5feb38f86219c6a14232273 Mon Sep 17 00:00:00 2001
From: wangyanfei01 <wangyanfei01@baidu.com>
Date: Tue, 6 Dec 2016 19:18:36 +0800
Subject: [PATCH] try to connect again if refused error found

---
 paddle/pserver/LightNetwork.cpp | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/paddle/pserver/LightNetwork.cpp b/paddle/pserver/LightNetwork.cpp
index 1830170a16..9b288604aa 100644
--- a/paddle/pserver/LightNetwork.cpp
+++ b/paddle/pserver/LightNetwork.cpp
@@ -18,6 +18,7 @@ limitations under the License. */
 #include <netinet/in.h>
 #include <netinet/tcp.h>
 #include <fcntl.h>
+#include <chrono>
 
 #include <arpa/inet.h>
 #include <sys/ioctl.h>
@@ -49,6 +50,10 @@ P_DEFINE_int32(sock_recv_buf_size,
                1024 * 1024 * 40,
                "restrict sock recv buff size");
 
+P_DEFINE_int32(connrefused_retries_second,
+               10,
+               "retry connrefused_retries_second if ECONNREFUSED occurs");
+
 namespace paddle {
 
 /**
@@ -382,8 +387,20 @@ void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) {
   setOption(sockfd);
 
   /// Now connect to the server
-  PCHECK(connect(sockfd, (sockaddr *)&serv_addr, sizeof(serv_addr)) >= 0)
-      << "ERROR connecting to " << serverAddr;
+  int retry_second = 0;
+  int error = 0;
+  do {
+    error = connect(sockfd, (sockaddr *)&serv_addr, sizeof(serv_addr));
+    if (error == ECONNREFUSED) {
+      LOG(WARNING) << "connection refused by pserver, try again!";
+      if (retry_second++ >= FLAGS_connrefused_retries_second) {
+        LOG(FATAL) << "connection refused by pserver, maybe pserver failed!";
+      }
+      std::this_thread::sleep_for(std::chrono::seconds(1));
+    } else {
+      PCHECK(error >= 0) << "ERROR connecting to " << serverAddr;
+    }
+  } while (error == ECONNREFUSED);
 
   channel_.reset(new SocketChannel(sockfd, serverAddr));
   tcpRdma_ = F_TCP;
-- 
GitLab