diff --git a/Makefile b/Makefile
index 45d072c..bdaab44 100644
--- a/Makefile
+++ b/Makefile
@@ -17,7 +17,7 @@ SRCS = $(sort $(wildcard src/*.c))
 OBJS = $(SRCS:.c=.o)
 LOBJS = src/core.o src/common.o src/libproxychains.o src/shm.o src/allocator_thread.o src/ip_type.o
 
-CFLAGS  += -Wall -O0 -g -std=c99 -D_GNU_SOURCE -pipe -DTHREAD_SAFE
+CFLAGS  += -Wall -O0 -g -std=c99 -D_GNU_SOURCE -pipe
 LDFLAGS = -shared -fPIC -Wl,--no-as-needed -ldl -lpthread
 INC     = 
 PIC     = -fPIC
diff --git a/src/core.c b/src/core.c
index c7ed7b2..0c5d159 100644
--- a/src/core.c
+++ b/src/core.c
@@ -34,10 +34,8 @@
 #include <sys/time.h>
 #include <stdarg.h>
 #include <assert.h>
-#ifdef THREAD_SAFE
-#include <pthread.h>
+#include "mutex.h"
 pthread_mutex_t hostdb_lock;
-#endif
 
 #include "core.h"
 #include "common.h"
@@ -692,6 +690,14 @@ int connect_proxy_chain(int sock, ip_type target_ip,
 	return -1;
 }
 
+void core_initialize(void) {
+	MUTEX_INIT(&hostdb_lock);
+}
+
+void core_unload(void) {
+	MUTEX_DESTROY(&hostdb_lock);
+}
+
 static void gethostbyname_data_setstring(struct gethostbyname_data* data, char* name) {
 	snprintf(data->addr_name, sizeof(data->addr_name), "%s", name);
 	data->hostent_space.h_name = data->addr_name;
diff --git a/src/core.h b/src/core.h
index d6c088d..9e895b6 100644
--- a/src/core.h
+++ b/src/core.h
@@ -28,11 +28,6 @@
 
 #include "ip_type.h"
 
-#ifdef THREAD_SAFE
-#include "mutex.h"
-extern pthread_mutex_t hostdb_lock;
-#endif
-
 /*error codes*/
 typedef enum {
 	SUCCESS=0,
@@ -121,6 +116,9 @@ void proxy_freeaddrinfo(struct addrinfo *res);
 
 void pc_stringfromipv4(unsigned char *ip_buf_4_bytes, char *outbuf_16_bytes);
 
+void core_initialize(void);
+void core_unload(void);
+
 #include "debug.h"
 
 #endif
diff --git a/src/libproxychains.c b/src/libproxychains.c
index 37c1f88..d293006 100644
--- a/src/libproxychains.c
+++ b/src/libproxychains.c
@@ -32,6 +32,7 @@
 #include <sys/socket.h>
 #include <fcntl.h>
 #include <dlfcn.h>
+#include <pthread.h>
 
 
 #include "core.h"
@@ -64,9 +65,8 @@ localaddr_arg localnet_addr[MAX_LOCALNET];
 size_t num_localnet_addr = 0;
 unsigned int remote_dns_subnet = 224;
 
-#ifdef THREAD_SAFE
 pthread_once_t init_once = PTHREAD_ONCE_INIT;
-#endif
+
 static int init_l = 0;
 
 static inline void get_chain_data(proxy_data * pd, unsigned int *proxy_count, chain_type * ct);
@@ -97,7 +97,7 @@ static void* load_sym(char* symname, void* proxyfunc) {
 
 static void do_init(void) {
 	srand(time(NULL));
-	MUTEX_INIT(&hostdb_lock);
+	core_initialize();
 	at_init();
 	
 	/* read the config file */
@@ -115,18 +115,23 @@ static void do_init(void) {
 	init_l = 1;
 }
 
+#if 0
+/* FIXME this is currently unused.
+ * it is not strictly needed.
+ * maybe let it be called by a gcc destructor, if that doesnt
+ * have negative consequences (e.g. when a child calles exit) */
+static void unload(void) {
+	at_close();
+	core_unload();
+}
+#endif
+
 static void init_lib_wrapper(const char* caller) {
 #ifndef DEBUG
 	(void) caller;
 #endif
-#ifndef THREAD_SAFE
-	if(init_l) return;
-	PDEBUG("%s called from %s\n", __FUNCTION__,  caller);
-	do_init();
-#else
 	if(!init_l) PDEBUG("%s called from %s\n", __FUNCTION__,  caller);
 	pthread_once(&init_once, do_init);
-#endif
 }
 
 /* if we use gcc >= 3, we can instruct the dynamic loader