diff --git a/pypy/module/cpyext/test/test_signature.py b/pypy/module/cpyext/test/test_signature.py
diff --git a/rpython/jit/metainterp/test/test_fficall.py b/rpython/jit/metainterp/test/test_fficall.py
@@ -1,4 +1,4 @@
-import py
+import py, pytest
 from _pytest.monkeypatch import monkeypatch
 import sys
 import ctypes, math
@@ -49,10 +49,13 @@ def _run(self, atypes, rtype, avalues, rvalue,
              expected_call_release_gil_i=1,
              expected_call_release_gil_f=0,
              expected_call_release_gil_n=0,
+             expected_call_may_force_i=0,
              expected_call_may_force_f=0,
+             expected_call_may_force_n=0,
              supports_floats=True,
              supports_longlong=False,
-             supports_singlefloats=False):
+             supports_singlefloats=False,
+             releasegil=True):
 
         cif_description = get_description(atypes, rtype)
 
@@ -75,7 +78,7 @@ def verify(*args):
         unroll_avalues = unrolling_iterable(avalues)
         BIG_ENDIAN = (sys.byteorder == 'big')
 
-        def fake_call_impl_any(cif_description, func_addr, exchange_buffer):
+        def fake_call_impl_any(cif_description, func_addr, exchange_buffer, releasegil=True):
             ofs = 16
             for avalue in unroll_avalues:
                 TYPE = rffi.CArray(lltype.typeOf(avalue))
@@ -117,7 +120,7 @@ def f(i):
                 rffi.cast(lltype.Ptr(TYPE), targetptr)[0] = avalue
                 targetptr = rffi.ptradd(targetptr, 16)
 
-            jit_ffi_call(cif_description, func_addr, exbuf)
+            jit_ffi_call(cif_description, func_addr, exbuf, releasegil=releasegil)
 
             if rvalue is None:
                 res = 654321
@@ -152,9 +155,9 @@ def matching_result(res, rvalue):
                 # longlong and floats are passed around as longlongs.
                 res = float2longlong(res)
             assert matching_result(res, rvalue)
-            self.check_operations_history(call_may_force_i=0,
+            self.check_operations_history(call_may_force_i=expected_call_may_force_i,
                             call_may_force_f=expected_call_may_force_f,
-                                          call_may_force_n=0,
+                            call_may_force_n=expected_call_may_force_n,
                             call_release_gil_i=expected_call_release_gil_i,
                             call_release_gil_f=expected_call_release_gil_f,
                             call_release_gil_n=expected_call_release_gil_n)
@@ -386,3 +389,9 @@ def test_calldescrof_dynamic_returning_none(self):
                                         expected_call_may_force_f=1)
         finally:
             LLGraphCPU.calldescrof_dynamic = old
+
+    @pytest.mark.xfail
+    def test_simple_call_int_dont_release_gil(self):
+        self._run([types.signed] * 2, types.signed, [456, 789], -42, releasegil=False,
+                  expected_call_release_gil_i=0, expected_call_may_force_i=1)
+
diff --git a/rpython/rlib/clibffi.py b/rpython/rlib/clibffi.py
@@ -344,6 +344,10 @@ def get_libc_name():
 c_ffi_call = external('ffi_call', [FFI_CIFP, rffi.VOIDP, rffi.VOIDP,
                                    VOIDPP], c_ffi_call_return_type,
                       save_err=rffi.RFFI_ERR_ALL | rffi.RFFI_ALT_ERRNO)
+c_ffi_call_no_release_gil = external('ffi_call', [FFI_CIFP, rffi.VOIDP, rffi.VOIDP,
+                                                  VOIDPP], c_ffi_call_return_type,
+                                     save_err=rffi.RFFI_ERR_ALL |
+                                     rffi.RFFI_ALT_ERRNO, releasegil=False)
 # Note: the RFFI_ALT_ERRNO flag matches the one in pyjitpl.direct_libffi_call
 CALLBACK_TP = rffi.CCallback([FFI_CIFP, rffi.VOIDP, rffi.VOIDPP, rffi.VOIDP],
                              lltype.Void)

diff --git a/rpython/rlib/jit_libffi.py b/rpython/rlib/jit_libffi.py
@@ -102,7 +102,7 @@ def jit_ffi_prep_cif(cif_description):
 ## have just read it from when called *in interpreter mode* only.
 
 
-def jit_ffi_call(cif_description, func_addr, exchange_buffer):
+def jit_ffi_call(cif_description, func_addr, exchange_buffer, releasegil=True):
     """Wrapper around ffi_call().  Must receive a CIF_DESCRIPTION_P that
     describes the layout of the 'exchange_buffer'.
 
@@ -131,7 +131,8 @@ def jit_ffi_call(cif_description, func_addr, exchange_buffer):
         #
         # Since call_release_gil is not generated, there is no need to
         # jit_ffi_save_result
-        jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer)
+        jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer,
+                              releasegil=releasegil)
 
 
 _short_sint_types = unrolling_iterable([rffi.SIGNEDCHAR, rffi.SHORT, rffi.INT])
@@ -231,7 +232,7 @@ def jit_ffi_call_impl_void(cif_description, func_addr, exchange_buffer):
     jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer)
     return None
 
-def jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer):
+def jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer, releasegil=True):
     """
     This is the function which actually calls libffi. All the rest is just
     infrastructure to convince the JIT to pass a typed result box to
@@ -243,9 +244,15 @@ def jit_ffi_call_impl_any(cif_description, func_addr, exchange_buffer):
         buffer_array[i] = data
     resultdata = rffi.ptradd(exchange_buffer,
                              cif_description.exchange_result)
-    clibffi.c_ffi_call(cif_description.cif, func_addr,
-                       rffi.cast(rffi.VOIDP, resultdata),
-                       buffer_array)
+    if releasegil:
+        clibffi.c_ffi_call(cif_description.cif, func_addr,
+                           rffi.cast(rffi.VOIDP, resultdata),
+                           buffer_array)
+    else:
+        clibffi.c_ffi_call_no_release_gil(cif_description.cif, func_addr,
+                                          rffi.cast(rffi.VOIDP, resultdata),
+                                          buffer_array)
+
 
 
 # ____________________________________________________________