diff --git a/pyopencl/c_wrapper/wrap_cl_core.h b/pyopencl/c_wrapper/wrap_cl_core.h
index 5e17b0de49a677c519609302524fa75df18e3b61..8a2cb02102f992490ab688d62ff6d8f68fc8f467 100644
--- a/pyopencl/c_wrapper/wrap_cl_core.h
+++ b/pyopencl/c_wrapper/wrap_cl_core.h
@@ -108,6 +108,7 @@ error *image__get_image_info(clobj_t image, cl_image_info param,
 error *event__get_profiling_info(clobj_t event, cl_profiling_info param,
                                  generic_info *out);
 error *event__wait(clobj_t event);
+error *wait_for_events(const clobj_t *_wait_for, uint32_t num_wait_for);
 // Nanny Event
 unsigned long nanny_event__get_ward(clobj_t evt);
 // enqueue_*
diff --git a/pyopencl/cffi_cl.py b/pyopencl/cffi_cl.py
index c6a3e5fb2ef7c1a2ef2982823d6508cc63489179..a2cef461be8107730ebbcee14299bc9badea9f35 100644
--- a/pyopencl/cffi_cl.py
+++ b/pyopencl/cffi_cl.py
@@ -696,7 +696,12 @@ class Event(_Common):
     def wait(self):
         _handle_error(_lib.event__wait(self.ptr))
 
+def wait_for_events(wait_for):
+    _handle_error(_lib.wait_for_events(*_c_obj_list(wait_for)))
+
 class NannyEvent(Event):
+    # TODO disable/handle write to buffer from bytes since the data may be moved
+    # by GC
     def get_ward(self):
         return _find_obj(_lib.nanny_event__get_ward(self.ptr))
 
diff --git a/src/c_wrapper/wrap_cl.cpp b/src/c_wrapper/wrap_cl.cpp
index fff574353f93d9ba1239c48af17f28a3b4644f98..7140d38ea6e3516b0d2b38d42fb47b19082107f2 100644
--- a/src/c_wrapper/wrap_cl.cpp
+++ b/src/c_wrapper/wrap_cl.cpp
@@ -2181,6 +2181,15 @@ event__wait(clobj_t evt)
         });
 }
 
+error*
+wait_for_events(const clobj_t *_wait_for, uint32_t num_wait_for)
+{
+    auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
+    return c_handle_error([&] {
+            pyopencl_call_guarded(clWaitForEvents,
+                                  num_wait_for, wait_for.get());
+        });
+}
 
 // Nanny Event
 unsigned long
@@ -2200,8 +2209,8 @@ enqueue_nd_range_kernel(clobj_t *_evt, clobj_t _queue, clobj_t _knl,
 {
     auto queue = static_cast<command_queue*>(_queue);
     auto knl = static_cast<kernel*>(_knl);
+    auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
     return c_handle_error([&] {
-            auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
             cl_event evt;
             retry_mem_error<void>([&] {
                     pyopencl_call_guarded(