diff --git a/src/backports/os.py b/src/backports/os.py
index 911897f..50aae86 100644
--- a/src/backports/os.py
+++ b/src/backports/os.py
@@ -94,13 +94,20 @@ def _chunks(b, indexes):
 
 def _fscodec():
     encoding = sys.getfilesystemencoding()
+
+    import codecs  # Use codecs.lookup() for name normalisation.
+    encoding = codecs.lookup(encoding).name
+    if encoding == 'ascii':
+        # force encoding to UTF-8 to use surrogate escape conversions
+        # when the encoding is bytes
+        encoding = 'utf-8'
+
     if encoding == 'mbcs':
         errors = 'strict'
     else:
         errors = 'surrogateescape'
 
     # XXX backport: Do we need to hack around Python 2's UTF-8 codec?
-    import codecs  # Use codecs.lookup() for name normalisation.
     _HACK_AROUND_PY2_UTF8 = (sys.version_info < (3,) and
                              codecs.lookup(encoding) == codecs.lookup('utf-8'))
 
diff --git a/tests/test_os.py b/tests/test_os.py
index 253c8d2..2ac029d 100644
--- a/tests/test_os.py
+++ b/tests/test_os.py
@@ -1,12 +1,13 @@
 """
-Partial backport of Python 3.5's Lib/test/test_os.py.
+Partial backport of Python 3.5's Lib/test/test_os.py with extra checks.
 """
 from __future__ import unicode_literals
-
-from backports import os
+from __future__ import absolute_import
 
 import unittest
 
+from backports import os
+
 
 class FSEncodingTests(unittest.TestCase):
     def test_nop(self):
@@ -14,10 +15,67 @@ def test_nop(self):
         self.assertEqual(os.fsdecode('abc\u0141'), 'abc\u0141')
 
     def test_identity(self):
-        # assert fsdecode(fsencode(x)) == x
         for fn in ('unicode\u0141', 'latin\xe9', 'ascii'):
             try:
                 bytesfn = os.fsencode(fn)
             except UnicodeEncodeError:
                 continue
             self.assertEqual(os.fsdecode(bytesfn), fn)
+
+    def test_non_utf(self):
+        test_data = [
+            b'non_ascii_12_\xe2\x82\xac',
+            b'non_ascii_2_\xc3\xa6',
+            b'non_ascii_5_\xcf\x86',
+            b'non_ascii_4_\xc5\x81',
+            b'non_utf8_decodable_\xff',
+            b'non_ascii_3_\xc4\xb0',
+            b'non_ascii_8_\xd8\x8c',
+            b'non_ascii_7_\xd7\x90',
+            b'non_utf8_decodable_3_\xed\xb4\x80',
+            b'non_ascii_-\xc3\xa0\xc3\xb2\xc9\x98\xc5\x81\xc4\x9f',
+            b'foo\xb1bar',
+            b'non_cp932_decodable_\xe7w\xf0',
+            b'non_ascii_11_\xc2\xa0',
+            b'non_utf8_decodable_2_\xed\xb2\x80',
+            b'non_ascii_9_\xd8\xaa',
+            b'non_cp12_decodable_\x81\x98',
+            b'non_ascii_10_\xe0\xb8\x81',
+            b'non_ascii_6_\xd0\x9a',
+            b'snow \xe2\x98\x83',
+        ]
+        for fn in test_data:
+            self.assertEqual(fn, os.fsencode(os.fsdecode(fn)))
+
+        fn = 'snow \u2603'
+        self.assertEqual(fn, os.fsdecode(os.fsencode(fn)))
+
+
+class FSEncodingTestsNoLocale(unittest.TestCase):
+    def test_non_utf_without_locale(self):
+        # this test will mimick a Linux without a proper locale and
+        # ascii fsencoding: it spawns this script to rerun the tests in
+        # a subprocess without a well define locale
+        import sys
+        import subprocess
+        import os as real_os
+
+        if 'linux' not in sys.platform:
+            return
+
+        cmd = sys.executable , real_os.path.abspath(__file__)
+
+        # unset langiuage env variales to ensure we have a plain POSIX
+        # locale and ascii FS encoding
+        env = {'LANG': b'', 'LANGUAGE': b''}
+        subprocess.check_output(cmd, env=env)
+
+
+def _suite():
+    suite = unittest.TestSuite()
+    suite.addTest(unittest.makeSuite(FSEncodingTests))
+    return suite
+
+
+if __name__ == '__main__':
+    unittest.main(defaultTest='_suite', verbosity=2)