Reject explicit redefinition of struct _IO_FILE in cdef()

doc/source/whatsnew.rst

@@ -11,8 +11,14 @@ v2.2.0.dev0
   (`#265`_).
 * The C source generated by ``cffi-gen-src`` and ``ffibuilder.emit_c_code()``
   no longer depends on the interpreter running the generator.
+* ``cdef()`` now raises a clean ``CDefError`` instead of crashing the whole
+  process with a fatal error when given an explicit definition of
+  ``struct _IO_FILE``, the name cffi uses internally for the opaque
+  ``FILE`` type. This could happen with an already-preprocessed system
+  header, which spells out the real struct. (`#281`_).
 
 .. _`#265`: https://github.com/python-cffi/cffi/pull/265
+.. _`#281`: https://github.com/python-cffi/cffi/pull/281
 
 v2.1.0
 ======

src/cffi/cparser.py

@@ -825,6 +825,23 @@ def _get_struct_union_enum_type(self, kind, type, name=None, nested=False):
         if type.decls is None:
             return tp
         #
+        if (kind == 'struct' and name is not None and
+                name == COMMON_TYPES['FILE'].name):
+            # cffi internally special-cases the C name used here (e.g.
+            # '_IO_FILE' on Linux) to mean the opaque FILE type: giving an
+            # explicit body for it collides with that special-casing and
+            # used to crash the process with a fatal error at runtime
+            # instead of failing cleanly here (issue #149).  This commonly
+            # happens when cdef() is fed an already-preprocessed system
+            # header, which spells out the real definition of this struct.
+            raise CDefError(
+                "'%s %s' is the struct name that cffi uses internally for "
+                "the opaque 'FILE' type; giving an explicit definition for "
+                "it in cdef() is not supported.  If this comes from a "
+                "preprocessed header, remove that struct definition (its "
+                "fields are not going to be used by cffi anyway)."
+                % (kind, name))
+        #
         if tp.fldnames is not None:
             raise CDefError("duplicate declaration of struct %s" % name)
         fldnames = []

testing/cffi0/test_parsing.py

@@ -382,6 +382,34 @@ def test_redefine_common_type():
     ffi = FFI()
     ffi.cdef("typedef bool (*fn_t)(bool, bool);")   # "bool," but within "( )"
 
+def test_explicit_struct_IO_FILE_rejected():
+    # Explicitly giving a body for the struct that cffi uses internally
+    # to back the opaque 'FILE' type used to build fine but then crash
+    # the whole process with a fatal error the first time the type was
+    # used (issue #149).  It's now rejected right away in cdef().  This
+    # is a realistic case: it happens when cdef() is fed a header that
+    # went through the C preprocessor, which spells out the real
+    # definition of 'struct _IO_FILE' coming from <stdio.h>.
+    ffi = FFI()
+    e = pytest.raises(CDefError, ffi.cdef, """
+        struct _IO_FILE {
+            int dummy;
+        };
+        typedef struct _IO_FILE FILE;
+    """)
+    assert 'struct _IO_FILE' in str(e.value)
+    assert 'FILE' in str(e.value)
+    # a forward declaration only (no body) is not ambiguous and is fine
+    ffi = FFI()
+    ffi.cdef("""
+        struct _IO_FILE;
+        typedef struct _IO_FILE FILE;
+        int fputs(const char *, FILE *);
+    """)
+    # plain use of FILE without redefining the struct still works
+    ffi = FFI()
+    ffi.cdef("int fputs(const char *, FILE *);")
+
 def test_bool():
     ffi = FFI()
     ffi.cdef("void f(bool);")