diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7b9b420b5..d890ef5cc 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -277,6 +277,7 @@ set(SOURCES
 	lib/parsers.c
 	lib/context.c
 	lib/sha-1.c
+	lib/alloc.c
 	)
 
 if (NOT LWS_WITHOUT_CLIENT)
diff --git a/lib/alloc.c b/lib/alloc.c
new file mode 100644
index 000000000..3bc5260ee
--- /dev/null
+++ b/lib/alloc.c
@@ -0,0 +1,30 @@
+#include "private-libwebsockets.h"
+
+static void *_realloc(void *ptr, size_t size)
+{
+	if (size)
+		return realloc(ptr, size);
+	else if (ptr)
+		free(ptr);
+	return NULL;
+}
+
+void *(*_lws_realloc)(void *ptr, size_t size) = _realloc;
+
+void *lws_realloc(void *ptr, size_t size)
+{
+	return _lws_realloc(ptr, size);
+}
+
+void *lws_zalloc(size_t size)
+{
+	void *ptr = _lws_realloc(NULL, size);
+	if (ptr)
+		memset(ptr, 0, size);
+	return ptr;
+}
+
+void lws_set_allocator(void *(*cb)(void *ptr, size_t size))
+{
+	_lws_realloc = cb;
+}
diff --git a/lib/client-handshake.c b/lib/client-handshake.c
index 17d77a5da..ca550bfcd 100644
--- a/lib/client-handshake.c
+++ b/lib/client-handshake.c
@@ -262,8 +262,8 @@ struct libwebsocket *libwebsocket_client_connect_2(
 	return wsi;
 
 oom4:
-	free(wsi->u.hdr.ah);
-	free(wsi);
+	lws_free(wsi->u.hdr.ah);
+	lws_free(wsi);
 	return NULL;
 
 failed:
@@ -305,11 +305,10 @@ libwebsocket_client_connect(struct libwebsocket_context *context,
 {
 	struct libwebsocket *wsi;
 
-	wsi = (struct libwebsocket *) malloc(sizeof(struct libwebsocket));
+	wsi = lws_zalloc(sizeof(struct libwebsocket));
 	if (wsi == NULL)
 		goto bail;
 
-	memset(wsi, 0, sizeof(*wsi));
 	wsi->sock = -1;
 
 	/* -1 means just use latest supported */
@@ -390,9 +389,9 @@ libwebsocket_client_connect(struct libwebsocket_context *context,
        return libwebsocket_client_connect_2(context, wsi);
 
 bail1:
-	free(wsi->u.hdr.ah);
+	lws_free(wsi->u.hdr.ah);
 bail:
-	free(wsi);
+	lws_free(wsi);
 
 	return NULL;
 }
diff --git a/lib/client.c b/lib/client.c
index 3c94c4fee..ef4f70e16 100755
--- a/lib/client.c
+++ b/lib/client.c
@@ -675,15 +675,12 @@ check_extensions:
 
 			wsi->active_extensions_user[
 				wsi->count_active_extensions] =
-					 malloc(ext->per_session_data_size);
+					 lws_zalloc(ext->per_session_data_size);
 			if (wsi->active_extensions_user[
 				wsi->count_active_extensions] == NULL) {
 				lwsl_err("Out of mem\n");
 				goto bail2;
 			}
-			memset(wsi->active_extensions_user[
-				wsi->count_active_extensions], 0,
-						    ext->per_session_data_size);
 			wsi->active_extensions[
 				  wsi->count_active_extensions] = ext;
 
@@ -744,8 +741,7 @@ check_accept:
 
 	/* free up his parsing allocations */
 
-	if (wsi->u.hdr.ah)
-		free(wsi->u.hdr.ah);
+	lws_free(wsi->u.hdr.ah);
 
 	lws_union_transition(wsi, LWS_CONNMODE_WS_CLIENT);
 	wsi->state = WSI_STATE_ESTABLISHED;
@@ -762,7 +758,7 @@ check_accept:
 	if (!n)
 		n = LWS_MAX_SOCKET_IO_BUF;
 	n += LWS_SEND_BUFFER_PRE_PADDING + LWS_SEND_BUFFER_POST_PADDING;
-	wsi->u.ws.rx_user_buffer = malloc(n);
+	wsi->u.ws.rx_user_buffer = lws_malloc(n);
 	if (!wsi->u.ws.rx_user_buffer) {
 		lwsl_err("Out of Mem allocating rx buffer %d\n", n);
 		goto bail2;
@@ -804,8 +800,7 @@ check_accept:
 	return 0;
 
 bail3:
-	free(wsi->u.ws.rx_user_buffer);
-	wsi->u.ws.rx_user_buffer = NULL;
+	lws_free2(wsi->u.ws.rx_user_buffer);
 	close_reason = LWS_CLOSE_STATUS_NOSTATUS;
 
 bail2:
@@ -818,8 +813,7 @@ bail2:
 
 	/* free up his parsing allocations */
 
-	if (wsi->u.hdr.ah)
-		free(wsi->u.hdr.ah);
+	lws_free(wsi->u.hdr.ah);
 
 	libwebsocket_close_and_free_session(context, wsi, close_reason);
 
diff --git a/lib/context.c b/lib/context.c
index 635bb1372..2ea9c60de 100644
--- a/lib/context.c
+++ b/lib/context.c
@@ -101,13 +101,11 @@ libwebsocket_create_context(struct lws_context_creation_info *info)
 	if (lws_plat_context_early_init())
 		return NULL;
 
-	context = (struct libwebsocket_context *)
-				malloc(sizeof(struct libwebsocket_context));
+	context = lws_zalloc(sizeof(struct libwebsocket_context));
 	if (!context) {
 		lwsl_err("No memory for websocket context\n");
 		return NULL;
 	}
-	memset(context, 0, sizeof(*context));
 
 	if (pid_daemon) {
 		context->started_with_parent = pid_daemon;
@@ -138,33 +136,29 @@ libwebsocket_create_context(struct lws_context_creation_info *info)
 					sizeof(struct libwebsocket *)) *
 							     context->max_fds));
 
-	context->fds = (struct libwebsocket_pollfd *)
-				malloc(sizeof(struct libwebsocket_pollfd) *
-							      context->max_fds);
+	context->fds = lws_zalloc(sizeof(struct libwebsocket_pollfd) *
+				  context->max_fds);
 	if (context->fds == NULL) {
 		lwsl_err("Unable to allocate fds array for %d connections\n",
 							      context->max_fds);
-		free(context);
+		lws_free(context);
 		return NULL;
 	}
 
-	context->lws_lookup = (struct libwebsocket **)
-		      malloc(sizeof(struct libwebsocket *) * context->max_fds);
+	context->lws_lookup = lws_zalloc(sizeof(struct libwebsocket *) * context->max_fds);
 	if (context->lws_lookup == NULL) {
 		lwsl_err(
 		  "Unable to allocate lws_lookup array for %d connections\n",
 							      context->max_fds);
-		free(context->fds);
-		free(context);
+		lws_free(context->fds);
+		lws_free(context);
 		return NULL;
 	}
-	memset(context->lws_lookup, 0, sizeof(struct libwebsocket *) *
-							context->max_fds);
 
 	if (lws_plat_init_fd_tables(context)) {
-		free(context->lws_lookup);
-		free(context->fds);
-		free(context);
+		lws_free(context->lws_lookup);
+		lws_free(context->fds);
+		lws_free(context);
 		return NULL;
 	}
 
@@ -334,12 +328,10 @@ libwebsocket_context_destroy(struct libwebsocket_context *context)
 
 	lws_ssl_context_destroy(context);
 
-	if (context->fds)
-		free(context->fds);
-	if (context->lws_lookup)
-		free(context->lws_lookup);
+	lws_free(context->fds);
+	lws_free(context->lws_lookup);
 
 	lws_plat_context_late_destroy(context);
 
-	free(context);
+	lws_free(context);
 }
diff --git a/lib/daemonize.c b/lib/daemonize.c
index 9e6131016..f5caf0b19 100644
--- a/lib/daemonize.c
+++ b/lib/daemonize.c
@@ -76,8 +76,7 @@ static void lws_daemon_closing(int sigact)
 	if (getpid() == pid_daemon)
 		if (lock_path) {
 			unlink(lock_path);
-			free(lock_path);
-			lock_path = NULL;
+			lws_free2(lock_path);
 		}
 
 	kill(getpid(), SIGKILL);
@@ -124,7 +123,7 @@ lws_daemonize(const char *_lock_path)
 	}
 
 	n = strlen(_lock_path) + 1;
-	lock_path = malloc(n);
+	lock_path = lws_malloc(n);
 	if (!lock_path) {
 		fprintf(stderr, "Out of mem in lws_daemonize\n");
 		return 1;
diff --git a/lib/extension-deflate-frame.c b/lib/extension-deflate-frame.c
index 6cf6ac016..13b2b8ded 100644
--- a/lib/extension-deflate-frame.c
+++ b/lib/extension-deflate-frame.c
@@ -54,16 +54,14 @@ int lws_extension_callback_deflate_frame(
 		conn->buf_out_length = sizeof(conn->buf_out);
 		conn->compressed_out = 0;
 		conn->buf_pre = NULL;
-		conn->buf_in = (unsigned char *)
-				malloc(LWS_SEND_BUFFER_PRE_PADDING +
-					       conn->buf_in_length +
-					       LWS_SEND_BUFFER_POST_PADDING);
+		conn->buf_in = lws_malloc(LWS_SEND_BUFFER_PRE_PADDING +
+					  conn->buf_in_length +
+					  LWS_SEND_BUFFER_POST_PADDING);
 		if (!conn->buf_in)
 			goto bail;
-		conn->buf_out = (unsigned char *)
-				malloc(LWS_SEND_BUFFER_PRE_PADDING +
-						conn->buf_out_length +
-						LWS_SEND_BUFFER_POST_PADDING);
+		conn->buf_out = lws_malloc(LWS_SEND_BUFFER_PRE_PADDING +
+					   conn->buf_out_length +
+					   LWS_SEND_BUFFER_POST_PADDING);
 		if (!conn->buf_out)
 			goto bail;
 		lwsl_ext("zlibs constructed\n");
@@ -75,10 +73,9 @@ bail:
 		return -1;
 
 	case LWS_EXT_CALLBACK_DESTROY:
-		if (conn->buf_pre)
-			free(conn->buf_pre);
-		free(conn->buf_in);
-		free(conn->buf_out);
+		lws_free(conn->buf_pre);
+		lws_free(conn->buf_in);
+		lws_free(conn->buf_out);
 		conn->buf_pre_used = 0;
 		conn->buf_pre_length = 0;
 		conn->buf_in_length = 0;
@@ -106,10 +103,8 @@ bail:
 
 			if (conn->buf_pre_length < total_payload) {
 				conn->buf_pre_length = total_payload;
-				if (conn->buf_pre)
-					free(conn->buf_pre);
-				conn->buf_pre =
-				    (unsigned char *)malloc(total_payload + 4);
+				lws_free(conn->buf_pre);
+				conn->buf_pre = lws_malloc(total_payload + 4);
 				if (!conn->buf_pre) {
 					lwsl_err("Out of memory\n");
 					return -1;
@@ -180,10 +175,10 @@ bail:
 						LWS_MAX_ZLIB_CONN_BUFFER);
 				return -1;
 			}
-			conn->buf_in = (unsigned char *)realloc(conn->buf_in,
-					LWS_SEND_BUFFER_PRE_PADDING +
-						conn->buf_in_length +
-						 LWS_SEND_BUFFER_POST_PADDING);
+			conn->buf_in = lws_realloc(conn->buf_in,
+						   LWS_SEND_BUFFER_PRE_PADDING +
+						   conn->buf_in_length +
+						   LWS_SEND_BUFFER_POST_PADDING);
 			if (!conn->buf_in) {
 				lwsl_err("Out of memory\n");
 				return -1;
@@ -243,11 +238,10 @@ bail:
 						LWS_MAX_ZLIB_CONN_BUFFER);
 				return -1;
 			}
-			conn->buf_out = (unsigned char *)realloc(
-					conn->buf_out,
-					LWS_SEND_BUFFER_PRE_PADDING +
-						  conn->buf_out_length +
-						  LWS_SEND_BUFFER_POST_PADDING);
+			conn->buf_out = lws_realloc(conn->buf_out,
+						    LWS_SEND_BUFFER_PRE_PADDING +
+						    conn->buf_out_length +
+						    LWS_SEND_BUFFER_POST_PADDING);
 			if (!conn->buf_out) {
 				lwsl_err("Out of memory\n");
 				return -1;
diff --git a/lib/getifaddrs.c b/lib/getifaddrs.c
index 1fa47cea4..da71583a9 100644
--- a/lib/getifaddrs.c
+++ b/lib/getifaddrs.c
@@ -58,6 +58,7 @@
 #endif
 
 #include "getifaddrs.h"
+#include "private-libwebsockets.h"
 
 static int
 getifaddrs2(struct ifaddrs **ifap,
@@ -85,7 +86,7 @@ getifaddrs2(struct ifaddrs **ifap,
 
 	buf_size = 8192;
 	for (;;) {
-		buf = calloc(1, buf_size);
+		buf = lws_zalloc(buf_size);
 		if (buf == NULL) {
 			ret = ENOMEM;
 			goto error_out;
@@ -107,7 +108,7 @@ getifaddrs2(struct ifaddrs **ifap,
 
 		if (ifconf.ifc_len < (int)buf_size)
 			break;
-		free(buf);
+		lws_free(buf);
 		buf_size *= 2;
 	}
 
@@ -137,12 +138,12 @@ getifaddrs2(struct ifaddrs **ifap,
 			goto error_out;
 		}
 
-		*end = malloc(sizeof(**end));
+		*end = lws_malloc(sizeof(**end));
 
 		(*end)->ifa_next = NULL;
 		(*end)->ifa_name = strdup(ifr->ifr_name);
 		(*end)->ifa_flags = ifreq.ifr_flags;
-		(*end)->ifa_addr = malloc(salen);
+		(*end)->ifa_addr = lws_malloc(salen);
 		memcpy((*end)->ifa_addr, sa, salen);
 		(*end)->ifa_netmask = NULL;
 
@@ -150,11 +151,12 @@ getifaddrs2(struct ifaddrs **ifap,
 		/* fix these when we actually need them */
 		if (ifreq.ifr_flags & IFF_BROADCAST) {
 			(*end)->ifa_broadaddr =
-				malloc(sizeof(ifr->ifr_broadaddr));
+				lws_malloc(sizeof(ifr->ifr_broadaddr));
 			memcpy((*end)->ifa_broadaddr, &ifr->ifr_broadaddr,
 						    sizeof(ifr->ifr_broadaddr));
 		} else if (ifreq.ifr_flags & IFF_POINTOPOINT) {
-			(*end)->ifa_dstaddr = malloc(sizeof(ifr->ifr_dstaddr));
+			(*end)->ifa_dstaddr =
+				lws_malloc(sizeof(ifr->ifr_dstaddr));
 			memcpy((*end)->ifa_dstaddr, &ifr->ifr_dstaddr,
 						      sizeof(ifr->ifr_dstaddr));
 		} else
@@ -169,12 +171,12 @@ getifaddrs2(struct ifaddrs **ifap,
 	}
 	*ifap = start;
 	close(fd);
-	free(buf);
+	lws_free(buf);
 	return 0;
 
 error_out:
 	close(fd);
-	free(buf);
+	lws_free(buf);
 	errno = ret;
 
 	return -1;
@@ -209,18 +211,14 @@ freeifaddrs(struct ifaddrs *ifp)
 	struct ifaddrs *p, *q;
 
 	for (p = ifp; p; ) {
-		free(p->ifa_name);
-		if (p->ifa_addr)
-			free(p->ifa_addr);
-		if (p->ifa_dstaddr)
-			free(p->ifa_dstaddr);
-		if (p->ifa_netmask)
-			free(p->ifa_netmask);
-		if (p->ifa_data)
-			free(p->ifa_data);
+		lws_free(p->ifa_name);
+		lws_free(p->ifa_addr);
+		lws_free(p->ifa_dstaddr);
+		lws_free(p->ifa_netmask);
+		lws_free(p->ifa_data);
 		q = p;
 		p = p->ifa_next;
-		free(q);
+		lws_free(q);
 	}
 }
 
diff --git a/lib/hpack.c b/lib/hpack.c
index 58d84b70a..3ac01b031 100644
--- a/lib/hpack.c
+++ b/lib/hpack.c
@@ -278,17 +278,16 @@ static int lws_hpack_add_dynamic_header(struct libwebsocket *wsi, int token, cha
 	dyn = wsi->u.http2.hpack_dyn_table;
 
 	if (!dyn) {
-		dyn = malloc(sizeof(*dyn));
+		dyn = lws_zalloc(sizeof(*dyn));
 		if (!dyn)
 			return 1;
-		memset(dyn, 0, sizeof(*dyn));
 		wsi->u.http2.hpack_dyn_table = dyn;
 		
-		dyn->args = malloc(1024);
+		dyn->args = lws_malloc(1024);
 		if (!dyn->args)
 			goto bail1;
 		dyn->args_length = 1024;
-		dyn->entries = malloc(sizeof(dyn->entries[0]) * 20);
+		dyn->entries = lws_malloc(sizeof(dyn->entries[0]) * 20);
 		if (!dyn->entries)
 			goto bail2;
 		dyn->num_entries = 20;
@@ -312,13 +311,13 @@ static int lws_hpack_add_dynamic_header(struct libwebsocket *wsi, int token, cha
 	dyn->next++;
 	
 	return 0;
-	
+
 bail2:
-	free(dyn->args);
+	lws_free(dyn->args);
 bail1:
-	free(dyn);
+	lws_free(dyn);
 	wsi->u.http2.hpack_dyn_table = NULL;
-		
+
 	return ret;
 }
 
@@ -681,4 +680,4 @@ int lws_add_http2_header_status(struct libwebsocket_context *context,
 		return 1;
 
 	return 0;
-}
\ No newline at end of file
+}
diff --git a/lib/libwebsockets.c b/lib/libwebsockets.c
index 6a9f466b9..d873306fa 100644
--- a/lib/libwebsockets.c
+++ b/lib/libwebsockets.c
@@ -208,35 +208,26 @@ just_kill_connection:
 	remove_wsi_socket_from_fds(context, wsi);
 
 	wsi->state = WSI_STATE_DEAD_SOCKET;
-	
-	if (wsi->rxflow_buffer) {
-		free(wsi->rxflow_buffer);
-		wsi->rxflow_buffer = NULL;
-	}
+
+	lws_free2(wsi->rxflow_buffer);
 
 	if (wsi->mode == LWS_CONNMODE_HTTP2_SERVING && wsi->u.hdr.ah) {
-		free(wsi->u.hdr.ah);
-		wsi->u.hdr.ah = NULL;
+		lws_free2(wsi->u.hdr.ah);
 	}
-	
+
 	if ((old_state == WSI_STATE_ESTABLISHED ||
 	     wsi->mode == LWS_CONNMODE_WS_SERVING ||
 	     wsi->mode == LWS_CONNMODE_WS_CLIENT)) {
 
-		if (wsi->u.ws.rx_user_buffer) {
-			free(wsi->u.ws.rx_user_buffer);
-			wsi->u.ws.rx_user_buffer = NULL;
-		}
+		lws_free2(wsi->u.ws.rx_user_buffer);
 
 		if (wsi->truncated_send_malloc) {
 			/* not going to be completed... nuke it */
-			free(wsi->truncated_send_malloc);
-			wsi->truncated_send_malloc = NULL;
+			lws_free2(wsi->truncated_send_malloc);
 			wsi->truncated_send_len = 0;
 		}
 		if (wsi->u.ws.ping_payload_buf) {
-			free(wsi->u.ws.ping_payload_buf);
-			wsi->u.ws.ping_payload_buf = NULL;
+			lws_free2(wsi->u.ws.ping_payload_buf);
 			wsi->u.ws.ping_payload_alloc = 0;
 			wsi->u.ws.ping_payload_len = 0;
 		}
@@ -265,7 +256,7 @@ just_kill_connection:
 		lwsl_warn("extension destruction failed\n");
 #ifndef LWS_NO_EXTENSIONS
 	for (n = 0; n < wsi->count_active_extensions; n++)
-		free(wsi->active_extensions_user[n]);
+		lws_free(wsi->active_extensions_user[n]);
 #endif
 	/*
 	 * inform all extensions in case they tracked this guy out of band
@@ -293,11 +284,11 @@ just_kill_connection:
 
 	if (wsi->protocol && wsi->protocol->per_session_data_size &&
 	    wsi->user_space && !wsi->user_space_externally_allocated)
-		free(wsi->user_space);
+		lws_free(wsi->user_space);
 
 	/* As a precaution, free the header table in case it lingered: */
 	lws_free_header_table(wsi);
-	free(wsi);
+	lws_free(wsi);
 }
 
 /**
@@ -710,14 +701,11 @@ libwebsocket_ensure_user_space(struct libwebsocket *wsi)
 	/* allocate the per-connection user memory (if any) */
 
 	if (wsi->protocol->per_session_data_size && !wsi->user_space) {
-		wsi->user_space = malloc(
-				  wsi->protocol->per_session_data_size);
+		wsi->user_space = lws_zalloc(wsi->protocol->per_session_data_size);
 		if (wsi->user_space  == NULL) {
 			lwsl_err("Out of memory for conn user space\n");
 			return 1;
 		}
-		memset(wsi->user_space, 0,
-					 wsi->protocol->per_session_data_size);
 	} else
 		lwsl_info("%s: %p protocol pss %u, user_space=%d\n", __func__, wsi, wsi->protocol->per_session_data_size, wsi->user_space);
 	return 0;
diff --git a/lib/libwebsockets.h b/lib/libwebsockets.h
index fa92f465c..ff0ca338f 100644
--- a/lib/libwebsockets.h
+++ b/lib/libwebsockets.h
@@ -160,6 +160,9 @@ LWS_VISIBLE LWS_EXTERN void lwsl_hexdump(void *buf, size_t len);
 /* you can call lws_get_peer_write_allowance */
 #define LWS_FEATURE_PROTOCOLS_HAS_PEER_WRITE_ALLOWANCE
 
+/* extra parameter introduced in 917f43ab821 */
+#define LWS_FEATURE_SERVE_HTTP_FILE_HAS_OTHER_HEADERS_LEN
+
 enum libwebsocket_context_options {
 	LWS_SERVER_OPTION_REQUIRE_VALID_OPENSSL_CLIENT_CERT = 2,
 	LWS_SERVER_OPTION_SKIP_SERVER_CANONICAL_NAME = 4,
@@ -1334,6 +1337,12 @@ libwebsocket_read(struct libwebsocket_context *context,
 LWS_VISIBLE LWS_EXTERN struct libwebsocket_extension *libwebsocket_get_internal_extensions();
 #endif
 
+/*
+ * custom allocator support
+ */
+LWS_VISIBLE LWS_EXTERN void
+lws_set_allocator(void *(*realloc)(void *ptr, size_t size));
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/lib/lws-plat-win.c b/lib/lws-plat-win.c
index 2c9c97f43..2503fc00f 100644
--- a/lib/lws-plat-win.c
+++ b/lib/lws-plat-win.c
@@ -193,17 +193,16 @@ lws_plat_drop_app_privileges(struct lws_context_creation_info *info)
 LWS_VISIBLE int
 lws_plat_init_fd_tables(struct libwebsocket_context *context)
 {
-	context->events = (WSAEVENT *)malloc(sizeof(WSAEVENT) *
-							(context->max_fds + 1));
+	context->events = lws_malloc(sizeof(WSAEVENT) * (context->max_fds + 1));
 	if (context->events == NULL) {
 		lwsl_err("Unable to allocate events array for %d connections\n",
 			context->max_fds);
 		return 1;
 	}
-	
+
 	context->fds_count = 0;
 	context->events[0] = WSACreateEvent();
-	
+
 	context->fd_random = 0;
 
 	return 0;
@@ -236,7 +235,7 @@ lws_plat_context_early_destroy(struct libwebsocket_context *context)
 {
 	if (context->events) {
 		WSACloseEvent(context->events[0]);
-		free(context->events);
+		lws_free(context->events);
 	}
 }
 
@@ -331,7 +330,7 @@ lws_plat_inet_ntop(int af, const void *src, char *dst, int cnt)
 	DWORD bufferlen = cnt;
 	BOOL ok = FALSE;
 
-	buffer = malloc(bufferlen);
+	buffer = lws_malloc(bufferlen);
 	if (!buffer) {
 		lwsl_err("Out of memory\n");
 		return NULL;
@@ -366,6 +365,6 @@ lws_plat_inet_ntop(int af, const void *src, char *dst, int cnt)
 			ok = FALSE;
 	}
 
-	free(buffer);
+	lws_free(buffer);
 	return ok ? dst : NULL;
 }
diff --git a/lib/output.c b/lib/output.c
index 846e3022c..b914f281c 100644
--- a/lib/output.c
+++ b/lib/output.c
@@ -195,11 +195,10 @@ handle_truncated_send:
 	 */
 	if (!wsi->truncated_send_malloc ||
 			real_len - n > wsi->truncated_send_allocation) {
-		if (wsi->truncated_send_malloc)
-			free(wsi->truncated_send_malloc);
+		lws_free(wsi->truncated_send_malloc);
 
 		wsi->truncated_send_allocation = real_len - n;
-		wsi->truncated_send_malloc = malloc(real_len - n);
+		wsi->truncated_send_malloc = lws_malloc(real_len - n);
 		if (!wsi->truncated_send_malloc) {
 			lwsl_err("truncated send: unable to malloc %d\n",
 							  real_len - n);
diff --git a/lib/parsers.c b/lib/parsers.c
index 0f0992724..2b4a900b0 100644
--- a/lib/parsers.c
+++ b/lib/parsers.c
@@ -63,7 +63,7 @@ int lws_allocate_header_table(struct libwebsocket *wsi)
 {
 	/* Be sure to free any existing header data to avoid mem leak: */
 	lws_free_header_table(wsi);
-	wsi->u.hdr.ah = malloc(sizeof(*wsi->u.hdr.ah));
+	wsi->u.hdr.ah = lws_malloc(sizeof(*wsi->u.hdr.ah));
 	if (wsi->u.hdr.ah == NULL) {
 		lwsl_err("Out of memory\n");
 		return -1;
@@ -77,11 +77,9 @@ int lws_allocate_header_table(struct libwebsocket *wsi)
 
 int lws_free_header_table(struct libwebsocket *wsi)
 {
-    if (wsi->u.hdr.ah) {
-        free(wsi->u.hdr.ah);
-        wsi->u.hdr.ah = NULL;
-    }
-    return 0;
+	lws_free2(wsi->u.hdr.ah);
+	wsi->u.hdr.ah = NULL;
+	return 0;
 };
 
 LWS_VISIBLE int lws_hdr_total_length(struct libwebsocket *wsi, enum lws_token_indexes h)
@@ -894,13 +892,13 @@ spill:
 			/* if existing buffer is too small, drop it */
 			if (wsi->u.ws.ping_payload_buf &&
 			    wsi->u.ws.ping_payload_alloc < wsi->u.ws.rx_user_buffer_head) {
-				free(wsi->u.ws.ping_payload_buf);
-				wsi->u.ws.ping_payload_buf = NULL;
+				lws_free2(wsi->u.ws.ping_payload_buf);
 			}
 
 			/* if no buffer, allocate it */
 			if (!wsi->u.ws.ping_payload_buf) {
-				wsi->u.ws.ping_payload_buf = malloc(wsi->u.ws.rx_user_buffer_head + LWS_SEND_BUFFER_PRE_PADDING);
+				wsi->u.ws.ping_payload_buf = lws_malloc(wsi->u.ws.rx_user_buffer_head
+									+ LWS_SEND_BUFFER_PRE_PADDING);
 				wsi->u.ws.ping_payload_alloc = wsi->u.ws.rx_user_buffer_head;
 			}
 			
diff --git a/lib/private-libwebsockets.h b/lib/private-libwebsockets.h
index f053ae769..c4550e916 100755
--- a/lib/private-libwebsockets.h
+++ b/lib/private-libwebsockets.h
@@ -1157,6 +1157,19 @@ lws_ssl_capable_write_no_ssl(struct libwebsocket *wsi, unsigned char *buf, int l
 #define lws_handshake_server(_a, _b, _c, _d) (0)
 #endif
 
+/*
+ * custom allocator
+ */
+LWS_EXTERN void*
+lws_realloc(void *ptr, size_t size);
+
+LWS_EXTERN void*
+lws_zalloc(size_t size);
+
+#define lws_malloc(S)	lws_realloc(NULL, S)
+#define lws_free(P)	lws_realloc(P, 0)
+#define lws_free2(P)	do { lws_realloc(P, 0); (P) = NULL; } while(0)
+
 /*
  * lws_plat_
  */
diff --git a/lib/server-handshake.c b/lib/server-handshake.c
index ddfa275e5..b554c6408 100644
--- a/lib/server-handshake.c
+++ b/lib/server-handshake.c
@@ -124,15 +124,12 @@ lws_extension_server_handshake(struct libwebsocket_context *context,
 
 			wsi->active_extensions_user[
 				wsi->count_active_extensions] =
-				     malloc(ext->per_session_data_size);
+				     lws_zalloc(ext->per_session_data_size);
 			if (wsi->active_extensions_user[
 			     wsi->count_active_extensions] == NULL) {
 				lwsl_err("Out of mem\n");
 				return 1;
 			}
-			memset(wsi->active_extensions_user[
-				wsi->count_active_extensions], 0,
-					    ext->per_session_data_size);
 
 			wsi->active_extensions[
 				  wsi->count_active_extensions] = ext;
@@ -154,7 +151,7 @@ lws_extension_server_handshake(struct libwebsocket_context *context,
 
 		n = 0;
 	}
-	
+
 	return 0;
 }
 #endif
diff --git a/lib/server.c b/lib/server.c
index cf266533a..0dbeb9d20 100644
--- a/lib/server.c
+++ b/lib/server.c
@@ -102,7 +102,7 @@ int lws_context_init_server(struct lws_context_creation_info *info,
 		compatible_close(sockfd);
 		return 1;
 	}
-	
+
 	if (getsockname(sockfd, (struct sockaddr *)&sin, &len) == -1)
 		lwsl_warn("getsockname: %s\n", strerror(LWS_ERRNO));
 	else
@@ -110,13 +110,12 @@ int lws_context_init_server(struct lws_context_creation_info *info,
 
 	context->listen_port = info->port;
 
-	wsi = (struct libwebsocket *)malloc(sizeof(struct libwebsocket));
+	wsi = lws_zalloc(sizeof(struct libwebsocket));
 	if (wsi == NULL) {
 		lwsl_err("Out of mem\n");
 		compatible_close(sockfd);
 		return 1;
 	}
-	memset(wsi, 0, sizeof(struct libwebsocket));
 	wsi->sock = sockfd;
 	wsi->mode = LWS_CONNMODE_SERVER_LISTENER;
 
@@ -128,7 +127,7 @@ int lws_context_init_server(struct lws_context_creation_info *info,
 
 	listen(sockfd, LWS_SOMAXCONN);
 	lwsl_notice(" Listening on port %d\n", info->port);
-	
+
 	return 0;
 }
 
@@ -296,10 +295,7 @@ got_uri:
 	}
 
 	/* now drop the header info we kept a pointer to */
-	if (wsi->u.http.ah)
-		free(wsi->u.http.ah);
-	/* not possible to continue to use past here */
-	wsi->u.http.ah = NULL;
+	lws_free2(wsi->u.http.ah);
 
 	if (n) {
 		lwsl_info("LWS_CALLBACK_HTTP closing\n");
@@ -322,11 +318,8 @@ got_uri:
 
 bail_nuke_ah:
 	/* drop the header info */
-	if (wsi->u.hdr.ah) {
-		free(wsi->u.hdr.ah);
-		wsi->u.hdr.ah = NULL;
-	}
-	
+	lws_free2(wsi->u.hdr.ah);
+
 	return 1;
 }
 
@@ -560,7 +553,7 @@ upgrade_ws:
 		if (!n)
 			n = LWS_MAX_SOCKET_IO_BUF;
 		n += LWS_SEND_BUFFER_PRE_PADDING + LWS_SEND_BUFFER_POST_PADDING;
-		wsi->u.ws.rx_user_buffer = malloc(n);
+		wsi->u.ws.rx_user_buffer = lws_malloc(n);
 		if (!wsi->u.ws.rx_user_buffer) {
 			lwsl_err("Out of Mem allocating rx buffer %d\n", n);
 			return 1;
@@ -589,13 +582,12 @@ libwebsocket_create_new_server_wsi(struct libwebsocket_context *context)
 {
 	struct libwebsocket *new_wsi;
 
-	new_wsi = (struct libwebsocket *)malloc(sizeof(struct libwebsocket));
+	new_wsi = lws_zalloc(sizeof(struct libwebsocket));
 	if (new_wsi == NULL) {
 		lwsl_err("Out of memory for new connection\n");
 		return NULL;
 	}
 
-	memset(new_wsi, 0, sizeof(struct libwebsocket));
 	new_wsi->pending_timeout = NO_PENDING_TIMEOUT;
 	new_wsi->rxflow_change_to = LWS_RXFLOW_ALLOW;
 
@@ -606,7 +598,7 @@ libwebsocket_create_new_server_wsi(struct libwebsocket_context *context)
 	new_wsi->hdr_parsing_completed = 0;
 
 	if (lws_allocate_header_table(new_wsi)) {
-		free(new_wsi);
+		lws_free(new_wsi);
 		return NULL;
 	}
 
diff --git a/lib/service.c b/lib/service.c
index ec3681b99..d06bba138 100644
--- a/lib/service.c
+++ b/lib/service.c
@@ -302,7 +302,7 @@ int lws_rxflow_cache(struct libwebsocket *wsi, unsigned char *buf, int n, int le
 
 	/* a new rxflow, buffer it and warn caller */
 	lwsl_info("new rxflow input buffer len %d\n", len - n);
-	wsi->rxflow_buffer = (unsigned char *)malloc(len - n);
+	wsi->rxflow_buffer = lws_malloc(len - n);
 	wsi->rxflow_len = len - n;
 	wsi->rxflow_pos = 0;
 	memcpy(wsi->rxflow_buffer, buf + n, len - n);
@@ -564,8 +564,7 @@ drain:
 		if (draining_flow && wsi->rxflow_buffer &&
 				 wsi->rxflow_pos == wsi->rxflow_len) {
 			lwsl_info("flow buffer: drained\n");
-			free(wsi->rxflow_buffer);
-			wsi->rxflow_buffer = NULL;
+			lws_free2(wsi->rxflow_buffer);
 			/* having drained the rxflow buffer, can rearm POLLIN */
 			_libwebsocket_rx_flow_control(wsi); /* n ignored, needed for NO_SERVER case */
 		}
diff --git a/lib/ssl.c b/lib/ssl.c
index 7cc565bc1..f9e88483e 100644
--- a/lib/ssl.c
+++ b/lib/ssl.c
@@ -480,7 +480,7 @@ lws_server_socket_service_ssl(struct libwebsocket_context *context,
 			    ERR_error_string(SSL_get_error(
 			    new_wsi->ssl, 0), NULL));
 			    libwebsockets_decode_ssl_error();
-			free(new_wsi);
+			lws_free(new_wsi);
 			compatible_close(accept_fd);
 			break;
 		}
diff --git a/test-server/test-echo.c b/test-server/test-echo.c
index 7a8b0023e..6cbb109b8 100644
--- a/test-server/test-echo.c
+++ b/test-server/test-echo.c
@@ -42,6 +42,7 @@
 #include "../lib/libwebsockets.h"
 
 static volatile int force_exit = 0;
+static int versa, state;
 
 #define MAX_ECHO_PAYLOAD 1400
 #define LOCAL_RESOURCE_PATH INSTALL_DATADIR"/libwebsockets-test-server"
@@ -67,6 +68,7 @@ callback_echo(struct libwebsocket_context *context,
 	/* when the callback is used for server operations --> */
 
 	case LWS_CALLBACK_SERVER_WRITEABLE:
+do_tx:
 		n = libwebsocket_write(wsi, &pss->buf[LWS_SEND_BUFFER_PRE_PADDING], pss->len, LWS_WRITE_TEXT);
 		if (n < 0) {
 			lwsl_err("ERROR %d writing to socket, hanging up\n", n);
@@ -79,6 +81,7 @@ callback_echo(struct libwebsocket_context *context,
 		break;
 
 	case LWS_CALLBACK_RECEIVE:
+do_rx:
 		if (len > MAX_ECHO_PAYLOAD) {
 			lwsl_err("Server received packet bigger than %u, hanging up\n", MAX_ECHO_PAYLOAD);
 			return 1;
@@ -92,16 +95,31 @@ callback_echo(struct libwebsocket_context *context,
 #ifndef LWS_NO_CLIENT
 	/* when the callback is used for client operations --> */
 
+	case LWS_CALLBACK_CLOSED:
+	case LWS_CALLBACK_CLIENT_CONNECTION_ERROR:
+		lwsl_info("closed\n");
+		state = 0;
+		break;
+
 	case LWS_CALLBACK_CLIENT_ESTABLISHED:
 		lwsl_notice("Client has connected\n");
 		pss->index = 0;
+		state = 2;
 		break;
 
 	case LWS_CALLBACK_CLIENT_RECEIVE:
+#ifndef LWS_NO_SERVER
+		if (versa)
+			goto do_rx;
+#endif
 		lwsl_notice("Client RX: %s", (char *)in);
 		break;
 
 	case LWS_CALLBACK_CLIENT_WRITEABLE:
+#ifndef LWS_NO_SERVER
+		if (versa)
+			goto do_tx;
+#endif
 		/* we will send our packet... */
 		pss->len = sprintf((char *)&pss->buf[LWS_SEND_BUFFER_PRE_PADDING], "hello from libwebsockets-test-echo client pid %d index %d\n", getpid(), pss->index++);
 		lwsl_notice("Client TX: %s", &pss->buf[LWS_SEND_BUFFER_PRE_PADDING]);
@@ -157,6 +175,8 @@ static struct option options[] = {
 	{ "ratems",	required_argument,	NULL, 'r' },
 #endif
 	{ "ssl",	no_argument,		NULL, 's' },
+	{ "versa",	no_argument,		NULL, 'v' },
+	{ "uri",	required_argument,	NULL, 'u' },
 	{ "passphrase", required_argument,	NULL, 'P' },
 	{ "interface",  required_argument,	NULL, 'i' },
 #ifndef LWS_NO_DAEMONIZE
@@ -183,6 +203,7 @@ int main(int argc, char **argv)
 	int listen_port = 80;
 	struct lws_context_creation_info info;
 	char passphrase[256];
+	char uri[256] = "/";
 #ifndef LWS_NO_CLIENT
 	char address[256], ads_port[256 + 30];
 	int rate_us = 250000;
@@ -206,7 +227,7 @@ int main(int argc, char **argv)
 #endif
 
 	while (n >= 0) {
-		n = getopt_long(argc, argv, "i:hsp:d:DC:k:P:"
+		n = getopt_long(argc, argv, "i:hsp:d:DC:k:P:vu:"
 #ifndef LWS_NO_CLIENT
 			"c:r:"
 #endif
@@ -228,7 +249,11 @@ int main(int argc, char **argv)
 			strncpy(ssl_key, optarg, sizeof(ssl_key));
 			ssl_key[sizeof(ssl_key) - 1] = '\0';
 			break;
-
+		case 'u':
+			strncpy(uri, optarg, sizeof(uri));
+			uri[sizeof(uri) - 1] = '\0';
+			break;
+			
 #ifndef LWS_NO_DAEMONIZE
 		case 'D':
 			daemonize = 1;
@@ -257,6 +282,9 @@ int main(int argc, char **argv)
 		case 'p':
 			port = atoi(optarg);
 			break;
+		case 'v':
+			versa = 1;
+			break;
 		case 'i':
 			strncpy(interface_name, optarg, sizeof interface_name);
 			interface_name[(sizeof interface_name) - 1] = '\0';
@@ -356,24 +384,7 @@ int main(int argc, char **argv)
 		return -1;
 	}
 
-#ifndef LWS_NO_CLIENT
-	if (client) {
-		lwsl_notice("Client connecting to %s:%u....\n", address, port);
-		/* we are in client mode */
-		
-		address[sizeof(address) - 1] = '\0';
-		sprintf(ads_port, "%s:%u\n", address, port & 65535);
-		
-		wsi = libwebsocket_client_connect(context, address,
-				port, use_ssl, "/", ads_port,
-				 "origin", NULL, -1);
-		if (!wsi) {
-			lwsl_err("Client failed to connect to %s:%u\n", address, port);
-			goto bail;
-		}
-		lwsl_notice("Client connected to %s:%u\n", address, port);
-	}
-#endif
+
 	signal(SIGINT, sighandler);
 
 	n = 0;
@@ -381,7 +392,24 @@ int main(int argc, char **argv)
 #ifndef LWS_NO_CLIENT
 		struct timeval tv;
 
-		if (client) {
+		if (client && !state) {
+			state = 1;
+			lwsl_notice("Client connecting to %s:%u....\n", address, port);
+			/* we are in client mode */
+		
+			address[sizeof(address) - 1] = '\0';
+			sprintf(ads_port, "%s:%u\n", address, port & 65535);
+		
+			wsi = libwebsocket_client_connect(context, address,
+				port, use_ssl, uri, ads_port,
+				 "origin", NULL, -1);
+			if (!wsi) {
+				lwsl_err("Client failed to connect to %s:%u\n", address, port);
+				goto bail;
+			}
+		}
+
+		if (client && !versa) {
 			gettimeofday(&tv, NULL);
 
 			if (((unsigned int)tv.tv_usec - oldus) > (unsigned int)rate_us) {
diff --git a/test-server/test-ping.c b/test-server/test-ping.c
index c04f28627..504af80da 100644
--- a/test-server/test-ping.c
+++ b/test-server/test-ping.c
@@ -190,7 +190,8 @@ callback_lws_mirror(struct libwebsocket_context * this,
 
 			if (!flood)
 				fprintf(stderr, "%d bytes from %s: req=%ld "
-				      "time=(unknown)\n", (int)len, address, l);
+				      "time=(unknown)\n", (int)len, address,
+					(long)l);
 			else
 				fprintf(stderr, "\b \b");
 
@@ -211,7 +212,7 @@ callback_lws_mirror(struct libwebsocket_context * this,
 
 		if (!flood)
 			fprintf(stderr, "%d bytes from %s: req=%ld "
-				"time=%lu.%lums\n", (int)len, address, l,
+				"time=%lu.%lums\n", (int)len, address, (long)l,
 			       (iv - psd->ringbuffer[n].issue_timestamp) / 1000,
 			((iv - psd->ringbuffer[n].issue_timestamp) / 100) % 10);
 		else