GitList
Browse code

Add workaround for a bug found in BSD kernels, which causes bogus error returned by the connect(2) system call in some rare conditions, resulting in inability to restart SEMS without restarting SER.

Should be no-op on !BSD systems.

Maxim Sobolev authored on 10/01/2005 17:31:19
Showing 2 changed files
Makefile.defs
History View file @ 746f3af828
... ... 
@@ -839,7 +839,7 @@ endif
839 839
840 840 
 ifeq ($(OS), freebsd)
841 841 
 	DEFS+=-DHAVE_SOCKADDR_SA_LEN -DHAVE_GETHOSTBYNAME2 -DHAVE_UNION_SEMUN \
842 
-		-DHAVE_SCHED_YIELD -DHAVE_MSGHDR_MSG_CONTROL
842 
+		-DHAVE_SCHED_YIELD -DHAVE_MSGHDR_MSG_CONTROL -DHAVE_CONNECT_ECONNRESET_BUG
843 843 
 	ifneq ($(found_lock_method), yes)
844 844 
 		DEFS+= -DUSE_PTHREAD_MUTEX  # try pthread sems
845 845 
 		found_lock_method=yes
... ... 
@@ -852,7 +852,7 @@ endif
852 852
853 853 
 ifeq ($(OS), openbsd)
854 854 
 	DEFS+=-DHAVE_SOCKADDR_SA_LEN  -DHAVE_GETHOSTBYNAME2 \
855 
-		-DHAVE_UNION_SEMUN -DHAVE_MSGHDR_MSG_CONTROL
855 
+		-DHAVE_UNION_SEMUN -DHAVE_MSGHDR_MSG_CONTROL -DHAVE_CONNECT_ECONNRESET_BUG
856 856 
 	ifneq ($(found_lock_method), yes)
857 857 
 		DEFS+= -DUSE_PTHREAD_MUTEX  # try pthread sems
858 858 
 		found_lock_method=yes
... ... 
@@ -881,7 +881,7 @@ endif   # if opensd
881 881
882 882 
 ifeq ($(OS), netbsd)
883 883 
 	DEFS+=-DHAVE_SOCKADDR_SA_LEN -DHAVE_GETHOSTBYNAME2 \
884 
-		-DHAVE_MSGHDR_MSG_CONTROL
884 
+		-DHAVE_MSGHDR_MSG_CONTROL -DHAVE_CONNECT_ECONNRESET_BUG
885 885 
 	ifneq ($(found_lock_method), yes)
886 886 
 		DEFS+= -DUSE_SYSV_SEM  # try pthread sems
887 887 
 		found_lock_method=yes
... ... 
@@ -895,7 +895,7 @@ ifeq ($(OS), darwin)
895 895 
 	DEFS+=-DHAVE_SOCKADDR_SA_LEN -DHAVE_GETHOSTBYNAME2 -DHAVE_UNION_SEMUN \
896 896 
 		-DHAVE_SCHED_YIELD -DHAVE_MSGHDR_MSG_CONTROL \
897 897 
 		-DUSE_ANON_MMAP \
898 
-		-DNDEBUG
898 
+		-DNDEBUG -DHAVE_CONNECT_ECONNRESET_BUG
899 899 
 	# -DNDEBUG used to turn off assert (assert wants to call
900 900 
 	# eprintf which doesn't seem to be defined in any shared lib
901 901 
 	ifneq ($(found_lock_method), yes)
modules/tm/t_fifo.c
History View file @ 746f3af828
... ... 
@@ -898,7 +898,7 @@ error:
898 898
899 899 
 static int write_to_unixsock(char* sockname, int cnt)
900 900 
 {
901 
-	int len;
901 
+	int len, e;
902 902 
 	struct sockaddr_un dest;
903 903
904 904 
 	if (!sockname) {
... ... 
@@ -921,12 +921,23 @@ static int write_to_unixsock(char* sockname, int cnt)
921 921 
 #ifdef HAVE_SOCKADDR_SA_LEN
922 922 
 	dest.sun_len = len;
923 923 
 #endif
924 
-
925 
-	if (connect(sock, (struct sockaddr*)&dest, SUN_LEN(&dest)) == -1) {
924 
+
925 
+	e = connect(sock, (struct sockaddr*)&dest, SUN_LEN(&dest));
926 
+#ifdef HAVE_CONNECT_ECONNRESET_BUG
927 
+	/*
928 
+	 * Workaround for a nasty bug in BSD kernels dated back
929 
+	 * to the Berkeley days, so that can be found in many modern
930 
+	 * BSD-derived kernels. Workaround should be pretty harmless since
931 
+	 * in normal conditions connect(2) can never return ECONNRESET.
932 
+	 */
933 
+	if ((e == -1) && (errno == ECONNRESET))
934 
+		e = 0;
935 
+#endif
936 
+	if (e == -1) {
926 937 
 		LOG(L_ERR, "write_to_unixsock: Error in connect: %s\n", strerror(errno));
927 938 
 		return -1;
928 939 
 	}
929 
-
940 
+
930 941 
 	if (tsend_dgram_ev(sock, (struct iovec*)lines_eol, 2 * cnt, tm_unix_tx_timeout * 1000) < 0) {
931 942 
 		LOG(L_ERR, "write_to_unixsock: writev failed: %s\n", strerror(errno));
932 943 
 		return -1;