Fix race in tfw_sock_srv_disconnect.

There was a race between `tfw_sock_srv_disconnect` and
`tfw_srv_conn_release` when last one is called from
`ss_conn_drop_guard_exit` when we process FIN from
remote peer. Connection can be released after we check
that connection refcount is not equal to TFW_CONN_DEATHCNT
and before we call `tfw_connection_close`. Later we
increment connection reference counter (for already
stopped connection, which is equal to zero) and put it
again. This leads to second connection release and
extra decrement of struct server reference counter.
We need to call `__tfw_connection_get_if_live` instead
of simple check that connection reference counter is
not equal to TFW_CONN_DEATHCNT before connection closing.

Closes #2047 #2054

fw/sock_srv.c

@@ -504,8 +504,9 @@ tfw_sock_srv_disconnect(TfwConn *conn)
 		 * restored already. If the connection is closed already, then
 		 * check its stop bit.
 		 */
-		if (atomic_read(&conn->refcnt) != TFW_CONN_DEATHCNT) {
+		if (__tfw_connection_get_if_live(conn)) {
 			TfwServer *srv = (TfwServer *)conn->peer;
+			int r = 0;
 
 			/*
 			 * We set TFW_CFG_B_DEL flag when we gracefully
@@ -521,10 +522,12 @@ tfw_sock_srv_disconnect(TfwConn *conn)
 			 */
 			if (test_bit(TFW_CFG_B_DEL, &srv->flags)) {
 				tfw_connection_abort(conn);
-				return 0;
+			} else {
+				r = tfw_connection_close(conn, true);
 			}
 
-			return tfw_connection_close(conn, true);
+			tfw_connection_put(conn);
+			return r;
 		}
 		/*
 		 * If stop flag is set, we can exit. Otherwise, continue waiting