Python script deadlocked when threads call get_torrent_status about libtorrent HOT 3 OPEN

--- a/bindings/python/src/session.cpp
+++ b/bindings/python/src/session.cpp
@@ -27,6 +27,8 @@
 #include <libtorrent/extensions/smart_ban.hpp>
 #include <libtorrent/extensions/ut_metadata.hpp>
 #include <libtorrent/extensions/ut_pex.hpp>
+#include <sstream>
+#include <sys/syscall.h>
 
 namespace boost
 {
@@ -521,6 +523,11 @@ namespace
 
     bool wrap_pred(object pred, torrent_status const& st)
     {
+        long int tid=syscall(SYS_gettid);
+        fprintf(stdout,"thread %-10ld gil %d\n",tid,PyGILState_Check());
+
+        PyEval_InitThreads();
+        lock_gil g; 
         return pred(st);
     }
 
@@ -531,12 +538,14 @@ namespace
         // libtorrent thread the python predicate will be freed from that
         // thread, which won't work
         auto wrapped_pred = std::bind(&wrap_pred, pred, std::placeholders::_1);
+        list ret;
+        Py_BEGIN_ALLOW_THREADS
         std::vector<torrent_status> torrents
             = s.get_torrent_status(std::ref(wrapped_pred), status_flags_t(flags));
 
-        list ret;
         for (std::vector<torrent_status>::iterator i = torrents.begin(); i != torrents.end(); ++i)
             ret.append(*i);
+        Py_END_ALLOW_THREADS
         return ret;
     }