Issue #13703: add a way to randomize the hash values of basic types (str, bytes, datetime)

in order to make algorithmic complexity attacks on (e.g.) web apps much more complicated.

The environment variable PYTHONHASHSEED and the new command line flag -R control this
behavior.

Doc/library/sys.rst

@@ -220,8 +220,12 @@ always available.
    :const:`ignore_environment`   :option:`-E`
    :const:`verbose`              :option:`-v`
    :const:`bytes_warning`        :option:`-b`
+   :const:`hash_randomization`   :option:`-R`
    ============================= =============================
 
+   .. versionadded:: 3.1.5
+      The ``hash_randomization`` attribute.
+
 
 .. data:: float_info
 

Doc/reference/datamodel.rst

@@ -1265,6 +1265,8 @@ Basic customization
    inheritance of :meth:`__hash__` will be blocked, just as if :attr:`__hash__`
    had been explicitly set to :const:`None`.
 
+   See also the :option:`-R` command-line option.
+
 
 .. method:: object.__bool__(self)
 

Doc/using/cmdline.rst

@@ -21,7 +21,7 @@ Command line
 
 When invoking Python, you may specify any of these options::
 
-    python [-bBdEhiOsSuvVWx?] [-c command | -m module-name | script | - ] [args]
+    python [-bBdEhiORsSuvVWx?] [-c command | -m module-name | script | - ] [args]
 
 The most common use case is, of course, a simple invocation of a script::
 
@@ -215,6 +215,29 @@ Miscellaneous options
    Discard docstrings in addition to the :option:`-O` optimizations.
 
 
+.. cmdoption:: -R
+
+   Turn on hash randomization, so that the :meth:`__hash__` values of str, bytes
+   and datetime objects are "salted" with an unpredictable random value.
+   Although they remain constant within an individual Python process, they are
+   not predictable between repeated invocations of Python.
+
+   This is intended to provide protection against a denial-of-service caused by
+   carefully-chosen inputs that exploit the worst case performance of a dict
+   insertion, O(n^2) complexity.  See
+   http://www.ocert.org/advisories/ocert-2011-003.html for details.
+
+   Changing hash values affects the order in which keys are retrieved from a
+   dict.  Although Python has never made guarantees about this ordering (and it
+   typically varies between 32-bit and 64-bit builds), enough real-world code
+   implicitly relies on this non-guaranteed behavior that the randomization is
+   disabled by default.
+
+   See also :envvar:`PYTHONHASHSEED`.
+
+   .. versionadded:: 3.1.5
+
+
 .. cmdoption:: -s
 
    Don't add user site directory to sys.path
@@ -314,6 +337,7 @@ Miscellaneous options
 
    .. note:: The line numbers in error messages will be off by one.
 
+
 Options you shouldn't use
 ~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -328,6 +352,7 @@ Options you shouldn't use
     Reserved for alternative implementations of Python to use for their own
     purposes.
 
+
 .. _using-on-envvars:
 
 Environment variables
@@ -435,6 +460,27 @@ These environment variables influence Python's behavior.
    import of source modules.
 
 
+.. envvar:: PYTHONHASHSEED
+
+   If this variable is set to ``random``, the effect is the same as specifying
+   the :option:`-R` option: a random value is used to seed the hashes of str,
+   bytes and datetime objects.
+
+   If :envvar:`PYTHONHASHSEED` is set to an integer value, it is used as a fixed
+   seed for generating the hash() of the types covered by the hash
+   randomization.
+
+   Its purpose is to allow repeatable hashing, such as for selftests for the
+   interpreter itself, or to allow a cluster of python processes to share hash
+   values.
+
+   The integer must be a decimal number in the range [0,4294967295].  Specifying
+   the value 0 will lead to the same hash values as when hash randomization is
+   disabled.
+
+   .. versionadded:: 3.1.5
+
+
 .. envvar:: PYTHONIOENCODING
 
    Overrides the encoding used for stdin/stdout/stderr, in the syntax

Include/object.h

@@ -473,6 +473,12 @@ PyAPI_FUNC(void) Py_ReprLeave(PyObject *);
 PyAPI_FUNC(long) _Py_HashDouble(double);
 PyAPI_FUNC(long) _Py_HashPointer(void*);
 
+typedef struct {
+    long prefix;
+    long suffix;
+} _Py_HashSecret_t;
+PyAPI_DATA(_Py_HashSecret_t) _Py_HashSecret;
+
 /* Helper for passing objects to printf and the like */
 #define PyObject_REPR(obj) _PyUnicode_AsString(PyObject_Repr(obj))
 

Include/pydebug.h

@@ -19,6 +19,7 @@ PyAPI_DATA(int) Py_DivisionWarningFlag;
 PyAPI_DATA(int) Py_DontWriteBytecodeFlag;
 PyAPI_DATA(int) Py_NoUserSiteDirectory;
 PyAPI_DATA(int) Py_UnbufferedStdioFlag;
+PyAPI_DATA(int) Py_HashRandomizationFlag;
 
 /* this is a wrapper around getenv() that pays attention to
    Py_IgnoreEnvironmentFlag.  It should be used for getting variables like

Include/pythonrun.h

@@ -174,6 +174,8 @@ typedef void (*PyOS_sighandler_t)(int);
 PyAPI_FUNC(PyOS_sighandler_t) PyOS_getsig(int);
 PyAPI_FUNC(PyOS_sighandler_t) PyOS_setsig(int, PyOS_sighandler_t);
 
+/* Random */
+PyAPI_FUNC(int) _PyOS_URandom (void *buffer, Py_ssize_t size);
 
 #ifdef __cplusplus
 }

Lib/json/__init__.py

@@ -31,7 +31,9 @@
 Compact encoding::
 
     >>> import json
-    >>> json.dumps([1,2,3,{'4': 5, '6': 7}], separators=(',', ':'))
+    >>> from collections import OrderedDict
+    >>> mydict = OrderedDict([('4', 5), ('6', 7)])
+    >>> json.dumps([1,2,3,mydict], separators=(',', ':'))
     '[1,2,3,{"4":5,"6":7}]'
 
 Pretty printing::

Lib/os.py

@@ -611,23 +611,6 @@ def _pickle_statvfs_result(sr):
 except NameError: # statvfs_result may not exist
     pass
 
-if not _exists("urandom"):
-    def urandom(n):
-        """urandom(n) -> str
-
-        Return a string of n random bytes suitable for cryptographic use.
-
-        """
-        try:
-            _urandomfd = open("/dev/urandom", O_RDONLY)
-        except (OSError, IOError):
-            raise NotImplementedError("/dev/urandom (or equivalent) not found")
-        bs = b""
-        while len(bs) < n:
-            bs += read(_urandomfd, n - len(bs))
-        close(_urandomfd)
-        return bs
-
 # Supply os.popen()
 def popen(cmd, mode="r", buffering=-1):
     if not isinstance(cmd, str):

Lib/test/mapping_tests.py

@@ -14,7 +14,7 @@ class BasicTestMappingProtocol(unittest.TestCase):
     def _reference(self):
         """Return a dictionary of values which are invariant by storage
         in the object under test."""
-        return {1:2, "key1":"value1", "key2":(1,2,3)}
+        return {"1": "2", "key1":"value1", "key2":(1,2,3)}
     def _empty_mapping(self):
         """Return an empty mapping object"""
         return self.type2test()

Lib/test/regrtest.py

@@ -428,6 +428,11 @@ def main(tests=None, testdir=None, verbose=0, quiet=False, generate=False,
         except ValueError:
             print("Couldn't find starting test (%s), using all tests" % start)
     if randomize:
+        hashseed = os.getenv('PYTHONHASHSEED')
+        if not hashseed:
+            os.environ['PYTHONHASHSEED'] = str(random_seed)
+            os.execv(sys.executable, [sys.executable] + sys.argv)
+            return
         random.seed(random_seed)
         print("Using random seed", random_seed)
         random.shuffle(tests)

Lib/test/script_helper.py

@@ -3,7 +3,6 @@
 
 import sys
 import os
-import re
 import os.path
 import tempfile
 import subprocess
@@ -19,11 +18,15 @@ def _assert_python(expected_success, *args, **env_vars):
     cmd_line = [sys.executable]
     if not env_vars:
         cmd_line.append('-E')
-    cmd_line.extend(args)
     # Need to preserve the original environment, for in-place testing of
     # shared library builds.
     env = os.environ.copy()
+    # But a special flag that can be set to override -- in this case, the
+    # caller is responsible to pass the full environment.
+    if env_vars.pop('__cleanenv', None):
+        env = {}
     env.update(env_vars)
+    cmd_line.extend(args)
     p = subprocess.Popen(cmd_line, stdin=subprocess.PIPE,
                          stdout=subprocess.PIPE, stderr=subprocess.PIPE,
                          env=env)

Lib/test/test_cmd_line.py

@@ -4,7 +4,6 @@
 
 import os
 import test.support, unittest
-import os
 import sys
 import subprocess
 
@@ -190,6 +189,22 @@ def test_large_PYTHONPATH(self):
             self.assertTrue(path1.encode('ascii') in stdout)
             self.assertTrue(path2.encode('ascii') in stdout)
 
+    def test_hash_randomization(self):
+        # Verify that -R enables hash randomization:
+        self.verify_valid_flag('-R')
+        hashes = []
+        for i in range(2):
+            code = 'print(hash("spam"))'
+            data, rc = self.start_python_and_exit_code('-R', '-c', code)
+            self.assertEqual(rc, 0)
+            hashes.append(data)
+        self.assertNotEqual(hashes[0], hashes[1])
+
+        # Verify that sys.flags contains hash_randomization
+        code = 'import sys; print("random is", sys.flags.hash_randomization)'
+        data, rc = self.start_python_and_exit_code('-R', '-c', code)
+        self.assertEqual(rc, 0)
+        self.assertIn(b'random is 1', data)
 
 def test_main():
     test.support.run_unittest(CmdLineTest)

Lib/test/test_descr.py

@@ -4300,8 +4300,18 @@ class C(metaclass=M):
 
     def test_repr(self):
         # Testing dict_proxy.__repr__
+        def sorted_dict_repr(repr_):
+            # Given the repr of a dict, sort the keys
+            assert repr_.startswith('{')
+            assert repr_.endswith('}')
+            kvs = repr_[1:-1].split(', ')
+            return '{' + ', '.join(sorted(kvs)) + '}'
         dict_ = {k: v for k, v in self.C.__dict__.items()}
-        self.assertEqual(repr(self.C.__dict__), 'dict_proxy({!r})'.format(dict_))
+        repr_ = repr(self.C.__dict__)
+        self.assert_(repr_.startswith('dict_proxy('))
+        self.assert_(repr_.endswith(')'))
+        self.assertEqual(sorted_dict_repr(repr_[len('dict_proxy('):-len(')')]),
+                         sorted_dict_repr('{!r}'.format(dict_)))
 
 
 class PTypesLongInitTest(unittest.TestCase):

Lib/test/test_hash.py

@@ -3,10 +3,16 @@
 #
 # Also test that hash implementations are inherited as expected
 
+import datetime
+import os
+import struct
 import unittest
 from test import support
+from test.script_helper import assert_python_ok
 from collections import Hashable
 
+IS_64BIT = (struct.calcsize('l') == 8)
+
 
 class HashEqualityTestCase(unittest.TestCase):
 
@@ -118,10 +124,92 @@ def test_hashes(self):
         for obj in self.hashes_to_check:
             self.assertEqual(hash(obj), _default_hash(obj))
 
+class HashRandomizationTests(unittest.TestCase):
+
+    # Each subclass should define a field "repr_", containing the repr() of
+    # an object to be tested
+
+    def get_hash_command(self, repr_):
+        return 'print(hash(%s))' % repr_
+
+    def get_hash(self, repr_, seed=None):
+        env = os.environ.copy()
+        env['__cleanenv'] = True  # signal to assert_python not to do a copy
+                                  # of os.environ on its own
+        if seed is not None:
+            env['PYTHONHASHSEED'] = str(seed)
+        else:
+            env.pop('PYTHONHASHSEED', None)
+        out = assert_python_ok(
+            '-c', self.get_hash_command(repr_),
+            **env)
+        stdout = out[1].strip()
+        return int(stdout)
+
+    def test_randomized_hash(self):
+        # two runs should return different hashes
+        run1 = self.get_hash(self.repr_, seed='random')
+        run2 = self.get_hash(self.repr_, seed='random')
+        self.assertNotEqual(run1, run2)
+
+class StringlikeHashRandomizationTests(HashRandomizationTests):
+    def test_null_hash(self):
+        # PYTHONHASHSEED=0 disables the randomized hash
+        if IS_64BIT:
+            known_hash_of_obj = 1453079729188098211
+        else:
+            known_hash_of_obj = -1600925533
+
+        # Randomization is disabled by default:
+        self.assertEqual(self.get_hash(self.repr_), known_hash_of_obj)
+
+        # It can also be disabled by setting the seed to 0:
+        self.assertEqual(self.get_hash(self.repr_, seed=0), known_hash_of_obj)
+
+    def test_fixed_hash(self):
+        # test a fixed seed for the randomized hash
+        # Note that all types share the same values:
+        if IS_64BIT:
+            h = -4410911502303878509
+        else:
+            h = -206076799
+        self.assertEqual(self.get_hash(self.repr_, seed=42), h)
+
+class StrHashRandomizationTests(StringlikeHashRandomizationTests):
+    repr_ = repr('abc')
+
+    def test_empty_string(self):
+        self.assertEqual(hash(""), 0)
+
+class BytesHashRandomizationTests(StringlikeHashRandomizationTests):
+    repr_ = repr(b'abc')
+
+    def test_empty_string(self):
+        self.assertEqual(hash(b""), 0)
+
+class DatetimeTests(HashRandomizationTests):
+    def get_hash_command(self, repr_):
+        return 'import datetime; print(hash(%s))' % repr_
+
+class DatetimeDateTests(DatetimeTests):
+    repr_ = repr(datetime.date(1066, 10, 14))
+
+class DatetimeDatetimeTests(DatetimeTests):
+    repr_ = repr(datetime.datetime(1, 2, 3, 4, 5, 6, 7))
+
+class DatetimeTimeTests(DatetimeTests):
+    repr_ = repr(datetime.time(0))
+
+
 def test_main():
     support.run_unittest(HashEqualityTestCase,
-                              HashInheritanceTestCase,
-                              HashBuiltinsTestCase)
+                         HashInheritanceTestCase,
+                         HashBuiltinsTestCase,
+                         StrHashRandomizationTests,
+                         BytesHashRandomizationTests,
+                         DatetimeDateTests,
+                         DatetimeDatetimeTests,
+                         DatetimeTimeTests)
 
 
 if __name__ == "__main__":

Lib/test/test_os.py

@@ -9,6 +9,7 @@
 import sys
 import shutil
 from test import support
+from test.script_helper import assert_python_ok
 
 # Detect whether we're on a Linux system that uses the (now outdated
 # and unmaintained) linuxthreads threading library.  There's an issue
@@ -574,14 +575,33 @@ def test_devnull(self):
         f.close()
 
 class URandomTests(unittest.TestCase):
-    def test_urandom(self):
-        try:
-            self.assertEqual(len(os.urandom(1)), 1)
-            self.assertEqual(len(os.urandom(10)), 10)
-            self.assertEqual(len(os.urandom(100)), 100)
-            self.assertEqual(len(os.urandom(1000)), 1000)
-        except NotImplementedError:
-            pass
+    def test_urandom_length(self):
+        self.assertEqual(len(os.urandom(0)), 0)
+        self.assertEqual(len(os.urandom(1)), 1)
+        self.assertEqual(len(os.urandom(10)), 10)
+        self.assertEqual(len(os.urandom(100)), 100)
+        self.assertEqual(len(os.urandom(1000)), 1000)
+
+    def test_urandom_value(self):
+        data1 = os.urandom(16)
+        data2 = os.urandom(16)
+        self.assertNotEqual(data1, data2)
+
+    def get_urandom_subprocess(self, count):
+        code = '\n'.join((
+            'import os, sys',
+            'data = os.urandom(%s)' % count,
+            'sys.stdout.buffer.write(data)',
+            'sys.stdout.buffer.flush()'))
+        out = assert_python_ok('-c', code)
+        stdout = out[1]
+        self.assertEqual(len(stdout), 16)
+        return stdout
+
+    def test_urandom_subprocess(self):
+        data1 = self.get_urandom_subprocess(16)
+        data2 = self.get_urandom_subprocess(16)
+        self.assertNotEqual(data1, data2)
 
 class ExecTests(unittest.TestCase):
     @unittest.skipIf(USING_LINUXTHREADS,

Lib/test/test_set.py

@@ -734,6 +734,17 @@ def test_repr(self):
         if self.repr is not None:
             self.assertEqual(repr(self.set), self.repr)
 
+    def check_repr_against_values(self):
+        text = repr(self.set)
+        self.assertTrue(text.startswith('{'))
+        self.assertTrue(text.endswith('}'))
+
+        result = text[1:-1].split(', ')
+        result.sort()
+        sorted_repr_values = [repr(value) for value in self.values]
+        sorted_repr_values.sort()
+        self.assertEqual(result, sorted_repr_values)
+
     def test_print(self):
         try:
             fo = open(support.TESTFN, "w")
@@ -892,7 +903,9 @@ def setUp(self):
         self.set    = set(self.values)
         self.dup    = set(self.values)
         self.length = 3
-        self.repr   = "{'a', 'c', 'b'}"
+
+    def test_repr(self):
+        self.check_repr_against_values()
 
 #------------------------------------------------------------------------------
 
@@ -903,7 +916,9 @@ def setUp(self):
         self.set    = set(self.values)
         self.dup    = set(self.values)
         self.length = 3
-        self.repr   = "{b'a', b'c', b'b'}"
+
+    def test_repr(self):
+        self.check_repr_against_values()
 
 #------------------------------------------------------------------------------
 
@@ -916,11 +931,13 @@ def setUp(self):
         self.set    = set(self.values)
         self.dup    = set(self.values)
         self.length = 4
-        self.repr   = "{'a', b'a', 'b', b'b'}"
 
     def tearDown(self):
         warnings.filters = self.warning_filters
 
+    def test_repr(self):
+        self.check_repr_against_values()
+
 #==============================================================================
 
 def baditer():

Lib/test/test_sys.py

@@ -446,7 +446,7 @@ def test_sys_flags(self):
         attrs = ("debug", "division_warning",
                  "inspect", "interactive", "optimize", "dont_write_bytecode",
                  "no_user_site", "no_site", "ignore_environment", "verbose",
-                 "bytes_warning")
+                 "bytes_warning", "hash_randomization")
         for attr in attrs:
             self.assertTrue(hasattr(sys.flags, attr), attr)
             self.assertEqual(type(getattr(sys.flags, attr)), int, attr)

Lib/test/test_urllib.py

@@ -12,6 +12,7 @@
 import sys
 import tempfile
 import warnings
+import collections
 
 def hexescape(char):
     """Escape char as RFC 2396 specifies"""
@@ -840,8 +841,9 @@ def test_nonstring_seq_values(self):
         self.assertEqual("a=1&a=2", urllib.parse.urlencode({"a": [1, 2]}, True))
         self.assertEqual("a=None&a=a",
                          urllib.parse.urlencode({"a": [None, "a"]}, True))
+        data = collections.OrderedDict([("a", 1), ("b", 1)])
         self.assertEqual("a=a&a=b",
-                         urllib.parse.urlencode({"a": {"a": 1, "b": 1}}, True))
+                         urllib.parse.urlencode({"a": data}, True))
 
     def test_urlencode_encoding(self):
         # ASCII encoding. Expect %3F with errors="replace'

Lib/tkinter/test/test_ttk/test_functions.py

@@ -143,7 +143,7 @@ def test_format_elemcreate(self):
             ('a', 'b', 'c')), ("test {a b} c", ()))
         # state spec and options
         self.assertEqual(ttk._format_elemcreate('image', False, 'test',
-            ('a', 'b'), a='x', b='y'), ("test a b", ("-a", "x", "-b", "y")))
+            ('a', 'b'), a='x'), ("test a b", ("-a", "x")))
         # format returned values as a tcl script
         # state spec with multiple states and an option with a multivalue
         self.assertEqual(ttk._format_elemcreate('image', True, 'test',

Makefile.pre.in

@@ -305,6 +305,7 @@ PYTHON_OBJS=	\
 		Python/pymath.o \
 		Python/pystate.o \
 		Python/pythonrun.o \
+		Python/random.o \
 		Python/structmember.o \
 		Python/symtable.o \
 		Python/sysmodule.o \

Misc/NEWS

@@ -10,6 +10,11 @@ What's New in Python 3.1.5?
 Core and Builtins
 -----------------
 
+- Issue #13703: oCERT-2011-003: add -R command-line option and PYTHONHASHSEED
+  environment variables, to provide an opt-in way to protect against denial of
+  service attacks due to hash collisions within the dict and set types.  Patch
+  by David Malcolm, based on work by Victor Stinner.
+
 Library
 -------
 

Misc/python.man

@@ -34,6 +34,9 @@ python \- an interpreted, interactive, object-oriented programming language
 .B \-OO
 ]
 [
+.B \-R
+]
+[
 .B -Q
 .I argument
 ]
@@ -145,6 +148,18 @@ to \fI.pyo\fP.  Given twice, causes docstrings to be discarded.
 .B \-OO
 Discard docstrings in addition to the \fB-O\fP optimizations.
 .TP
+.B \-R
+Turn on "hash randomization", so that the hash() values of str, bytes and
+datetime objects are "salted" with an unpredictable pseudo-random value.
+Although they remain constant within an individual Python process, they are
+not predictable between repeated invocations of Python.
+.IP
+This is intended to provide protection against a denial of service
+caused by carefully-chosen inputs that exploit the worst case performance
+of a dict insertion, O(n^2) complexity.  See
+http://www.ocert.org/advisories/ocert-2011-003.html
+for details.
+.TP
 .BI "\-Q " argument
 Division control; see PEP 238.  The argument must be one of "old" (the
 default, int/int and long/long return an int or long), "new" (new
@@ -403,6 +418,20 @@ the \fB\-u\fP option.
 If this is set to a non-empty string it is equivalent to specifying
 the \fB\-v\fP option. If set to an integer, it is equivalent to
 specifying \fB\-v\fP multiple times. 
+.IP PYTHONHASHSEED
+If this variable is set to "random", the effect is the same as specifying
+the \fB-R\fP option: a random value is used to seed the hashes of str,
+bytes and datetime objects.
+
+If PYTHONHASHSEED is set to an integer value, it is used as a fixed seed for
+generating the hash() of the types covered by the hash randomization.  Its
+purpose is to allow repeatable hashing, such as for selftests for the
+interpreter itself, or to allow a cluster of python processes to share hash
+values.
+
+The integer must be a decimal number in the range [0,4294967295].  Specifying
+the value 0 will lead to the same hash values as when hash randomization is
+disabled.
 .SH AUTHOR
 The Python Software Foundation: http://www.python.org/psf
 .SH INTERNET RESOURCES

Modules/datetimemodule.c

@@ -2566,10 +2566,12 @@ generic_hash(unsigned char *data, int len)
     register long x;
 
     p = (unsigned char *) data;
-    x = *p << 7;
+    x = _Py_HashSecret.prefix;
+    x ^= *p << 7;
     while (--len >= 0)
         x = (1000003*x) ^ *p++;
     x ^= len;
+    x ^= _Py_HashSecret.suffix;
     if (x == -1)
         x = -2;
 

Modules/main.c

@@ -47,7 +47,7 @@ static wchar_t **orig_argv;
 static int  orig_argc;
 
 /* command line options */
-#define BASE_OPTS L"bBc:dEhiJm:OsStuvVW:xX?"
+#define BASE_OPTS L"bBc:dEhiJm:ORsStuvVW:xX?"
 
 #define PROGRAM_OPTS BASE_OPTS
 
@@ -72,6 +72,9 @@ static char *usage_2 = "\
 -m mod : run library module as a script (terminates option list)\n\
 -O     : optimize generated bytecode slightly; also PYTHONOPTIMIZE=x\n\
 -OO    : remove doc-strings in addition to the -O optimizations\n\
+-R     : use a pseudo-random salt to make hash() values of various types be\n\
+         unpredictable between separate invocations of the interpreter, as\n\
+         a defence against denial-of-service attacks\n\
 -s     : don't add user site directory to sys.path; also PYTHONNOUSERSITE\n\
 -S     : don't imply 'import site' on initialization\n\
 ";
@@ -99,6 +102,12 @@ PYTHONHOME   : alternate <prefix> directory (or <prefix>%c<exec_prefix>).\n\
 PYTHONCASEOK : ignore case in 'import' statements (Windows).\n\
 PYTHONIOENCODING: Encoding[:errors] used for stdin/stdout/stderr.\n\
 ";
+static char *usage_6 = "\
+PYTHONHASHSEED: if this variable is set to ``random``, the effect is the same \n\
+   as specifying the :option:`-R` option: a random value is used to seed the\n\
+   hashes of str, bytes and datetime objects.  It can also be set to an integer\n\
+   in the range [0,4294967295] to get hash values with a predictable seed.\n\
+";
 
 #ifndef MS_WINDOWS
 static FILE*
@@ -136,6 +145,7 @@ usage(int exitcode, wchar_t* program)
         fputs(usage_3, f);
         fprintf(f, usage_4, DELIM);
         fprintf(f, usage_5, DELIM, PYTHONHOMEHELP);
+        fputs(usage_6, f);
     }
 #if defined(__VMS)
     if (exitcode == 0) {
@@ -373,6 +383,10 @@ Py_Main(int argc, wchar_t **argv)
             PySys_AddWarnOption(_PyOS_optarg);
             break;
 
+        case 'R':
+            Py_HashRandomizationFlag++;
+            break;
+
         /* This space reserved for other options */
 
         default:

Modules/posixmodule.c

@@ -4022,7 +4022,7 @@ posix_getgroups(PyObject *self, PyObject *noargs)
 #endif
     gid_t grouplist[MAX_GROUPS];
 
-    /* On MacOSX getgroups(2) can return more than MAX_GROUPS results 
+    /* On MacOSX getgroups(2) can return more than MAX_GROUPS results
      * This is a helper variable to store the intermediate result when
      * that happens.
      *
@@ -6942,82 +6942,6 @@ posix_getloadavg(PyObject *self, PyObject *noargs)
 }
 #endif
 
-#ifdef MS_WINDOWS
-
-PyDoc_STRVAR(win32_urandom__doc__,
-"urandom(n) -> str\n\n\
-Return n random bytes suitable for cryptographic use.");
-
-typedef BOOL (WINAPI *CRYPTACQUIRECONTEXTA)(HCRYPTPROV *phProv,\
-              LPCSTR pszContainer, LPCSTR pszProvider, DWORD dwProvType,\
-              DWORD dwFlags );
-typedef BOOL (WINAPI *CRYPTGENRANDOM)(HCRYPTPROV hProv, DWORD dwLen,\
-              BYTE *pbBuffer );
-
-static CRYPTGENRANDOM pCryptGenRandom = NULL;
-/* This handle is never explicitly released. Instead, the operating
-   system will release it when the process terminates. */
-static HCRYPTPROV hCryptProv = 0;
-
-static PyObject*
-win32_urandom(PyObject *self, PyObject *args)
-{
-    int howMany;
-    PyObject* result;
-
-    /* Read arguments */
-    if (! PyArg_ParseTuple(args, "i:urandom", &howMany))
-        return NULL;
-    if (howMany < 0)
-        return PyErr_Format(PyExc_ValueError,
-                            "negative argument not allowed");
-
-    if (hCryptProv == 0) {
-        HINSTANCE hAdvAPI32 = NULL;
-        CRYPTACQUIRECONTEXTA pCryptAcquireContext = NULL;
-
-        /* Obtain handle to the DLL containing CryptoAPI
-           This should not fail         */
-        hAdvAPI32 = GetModuleHandle("advapi32.dll");
-        if(hAdvAPI32 == NULL)
-            return win32_error("GetModuleHandle", NULL);
-
-        /* Obtain pointers to the CryptoAPI functions
-           This will fail on some early versions of Win95 */
-        pCryptAcquireContext = (CRYPTACQUIRECONTEXTA)GetProcAddress(
-                                        hAdvAPI32,
-                                        "CryptAcquireContextA");
-        if (pCryptAcquireContext == NULL)
-            return PyErr_Format(PyExc_NotImplementedError,
-                                "CryptAcquireContextA not found");
-
-        pCryptGenRandom = (CRYPTGENRANDOM)GetProcAddress(
-                                        hAdvAPI32, "CryptGenRandom");
-        if (pCryptGenRandom == NULL)
-            return PyErr_Format(PyExc_NotImplementedError,
-                                "CryptGenRandom not found");
-
-        /* Acquire context */
-        if (! pCryptAcquireContext(&hCryptProv, NULL, NULL,
-                                   PROV_RSA_FULL, CRYPT_VERIFYCONTEXT))
-            return win32_error("CryptAcquireContext", NULL);
-    }
-
-    /* Allocate bytes */
-    result = PyBytes_FromStringAndSize(NULL, howMany);
-    if (result != NULL) {
-        /* Get random data */
-        memset(PyBytes_AS_STRING(result), 0, howMany); /* zero seed */
-        if (! pCryptGenRandom(hCryptProv, howMany, (unsigned char*)
-                              PyBytes_AS_STRING(result))) {
-            Py_DECREF(result);
-            return win32_error("CryptGenRandom", NULL);
-        }
-    }
-    return result;
-}
-#endif
-
 PyDoc_STRVAR(device_encoding__doc__,
 "device_encoding(fd) -> str\n\n\
 Return a string describing the encoding of the device\n\
@@ -7055,41 +6979,35 @@ device_encoding(PyObject *self, PyObject *args)
     return Py_None;
 }
 
-#ifdef __VMS
-/* Use openssl random routine */
-#include <openssl/rand.h>
-PyDoc_STRVAR(vms_urandom__doc__,
+PyDoc_STRVAR(posix_urandom__doc__,
 "urandom(n) -> str\n\n\
 Return n random bytes suitable for cryptographic use.");
 
-static PyObject*
-vms_urandom(PyObject *self, PyObject *args)
+static PyObject *
+posix_urandom(PyObject *self, PyObject *args)
 {
-    int howMany;
-    PyObject* result;
+    Py_ssize_t size;
+    PyObject *result;
+    int ret;
 
-    /* Read arguments */
-    if (! PyArg_ParseTuple(args, "i:urandom", &howMany))
+     /* Read arguments */
+    if (!PyArg_ParseTuple(args, "n:urandom", &size))
         return NULL;
-    if (howMany < 0)
+    if (size < 0)
         return PyErr_Format(PyExc_ValueError,
                             "negative argument not allowed");
+    result = PyBytes_FromStringAndSize(NULL, size);
+    if (result == NULL)
+        return NULL;
 
-    /* Allocate bytes */
-    result = PyBytes_FromStringAndSize(NULL, howMany);
-    if (result != NULL) {
-        /* Get random data */
-        if (RAND_pseudo_bytes((unsigned char*)
-                              PyBytes_AS_STRING(result),
-                              howMany) < 0) {
-            Py_DECREF(result);
-            return PyErr_Format(PyExc_ValueError,
-                                "RAND_pseudo_bytes");
-        }
+    ret = _PyOS_URandom(PyBytes_AS_STRING(result),
+                        PyBytes_GET_SIZE(result));
+    if (ret == -1) {
+        Py_DECREF(result);
+        return NULL;
     }
     return result;
 }
-#endif
 
 static PyMethodDef posix_methods[] = {
     {"access",          posix_access, METH_VARARGS, posix_access__doc__},
@@ -7374,12 +7292,7 @@ static PyMethodDef posix_methods[] = {
 #ifdef HAVE_GETLOADAVG
     {"getloadavg",      posix_getloadavg, METH_NOARGS, posix_getloadavg__doc__},
 #endif
- #ifdef MS_WINDOWS
-    {"urandom", win32_urandom, METH_VARARGS, win32_urandom__doc__},
- #endif
- #ifdef __VMS
-    {"urandom", vms_urandom, METH_VARARGS, vms_urandom__doc__},
- #endif
+    {"urandom",         posix_urandom,   METH_VARARGS, posix_urandom__doc__},
     {NULL,              NULL}            /* Sentinel */
 };
 

Objects/bytesobject.c

@@ -899,11 +899,21 @@ bytes_hash(PyBytesObject *a)
     if (a->ob_shash != -1)
         return a->ob_shash;
     len = Py_SIZE(a);
+    /*
+      We make the hash of the empty string be 0, rather than using
+      (prefix ^ suffix), since this slightly obfuscates the hash secret
+    */
+    if (len == 0) {
+        a->ob_shash = 0;
+        return 0;
+    }
     p = (unsigned char *) a->ob_sval;
-    x = *p << 7;
+    x = _Py_HashSecret.prefix;
+    x ^= *p << 7;
     while (--len >= 0)
         x = (1000003*x) ^ *p++;
     x ^= Py_SIZE(a);
+    x ^= _Py_HashSecret.suffix;
     if (x == -1)
         x = -2;
     a->ob_shash = x;

Objects/object.c

@@ -712,6 +712,8 @@ PyObject_HashNotImplemented(PyObject *v)
     return -1;
 }
 
+_Py_HashSecret_t _Py_HashSecret;
+
 long
 PyObject_Hash(PyObject *v)
 {

Objects/unicodeobject.c

@@ -7344,11 +7344,21 @@ unicode_hash(PyUnicodeObject *self)
     if (self->hash != -1)
         return self->hash;
     len = Py_SIZE(self);
+    /*
+      We make the hash of the empty string be 0, rather than using
+      (prefix ^ suffix), since this slightly obfuscates the hash secret
+    */
+    if (len == 0) {
+        self->hash = 0;
+        return 0;
+    }
     p = self->str;
-    x = *p << 7;
+    x = _Py_HashSecret.prefix;
+    x ^= *p << 7;
     while (--len >= 0)
         x = (1000003*x) ^ *p++;
     x ^= Py_SIZE(self);
+    x ^= _Py_HashSecret.suffix;
     if (x == -1)
         x = -2;
     self->hash = x;

PCbuild/pythoncore.vcproj

@@ -1778,6 +1778,10 @@
 				RelativePath="..\Python\pythonrun.c"
 				>
 			</File>
+ 			<File
+				RelativePath="..\Python\random.c"
+				>
+			</File>
 			<File
 				RelativePath="..\Python\structmember.c"
 				>

Python/pythonrun.c

@@ -73,6 +73,7 @@ extern void _PyUnicode_Init(void);
 extern void _PyUnicode_Fini(void);
 extern int _PyLong_Init(void);
 extern void PyLong_Fini(void);
+extern void _PyRandom_Init(void);
 
 #ifdef WITH_THREAD
 extern void _PyGILState_Init(PyInterpreterState *, PyThreadState *);
@@ -91,6 +92,7 @@ int Py_FrozenFlag; /* Needed by getpath.c */
 int Py_IgnoreEnvironmentFlag; /* e.g. PYTHONPATH, PYTHONHOME */
 int Py_NoUserSiteDirectory = 0; /* for -s and site.py */
 int Py_UnbufferedStdioFlag = 0; /* Unbuffered binary std{in,out,err} */
+int Py_HashRandomizationFlag = 0; /* for -R and PYTHONHASHSEED */
 
 /* PyModule_GetWarningsModule is no longer necessary as of 2.6
 since _warnings is builtin.  This API should not be used. */
@@ -195,6 +197,12 @@ Py_InitializeEx(int install_sigs)
         Py_OptimizeFlag = add_flag(Py_OptimizeFlag, p);
     if ((p = Py_GETENV("PYTHONDONTWRITEBYTECODE")) && *p != '\0')
         Py_DontWriteBytecodeFlag = add_flag(Py_DontWriteBytecodeFlag, p);
+    /* The variable is only tested for existence here; _PyRandom_Init will
+       check its value further. */
+    if ((p = Py_GETENV("PYTHONHASHSEED")) && *p != '\0')
+        Py_HashRandomizationFlag = add_flag(Py_HashRandomizationFlag, p);
+
+    _PyRandom_Init();
 
     interp = PyInterpreterState_New();
     if (interp == NULL)

Python/random.c

@@ -0,0 +1,302 @@
+#include "Python.h"
+#ifdef MS_WINDOWS
+#include <windows.h>
+#else
+#include <fcntl.h>
+#endif
+
+static int random_initialized = 0;
+
+#ifdef MS_WINDOWS
+typedef BOOL (WINAPI *CRYPTACQUIRECONTEXTA)(HCRYPTPROV *phProv,\
+              LPCSTR pszContainer, LPCSTR pszProvider, DWORD dwProvType,\
+              DWORD dwFlags );
+typedef BOOL (WINAPI *CRYPTGENRANDOM)(HCRYPTPROV hProv, DWORD dwLen,\
+              BYTE *pbBuffer );
+
+static CRYPTGENRANDOM pCryptGenRandom = NULL;
+/* This handle is never explicitly released. Instead, the operating
+   system will release it when the process terminates. */
+static HCRYPTPROV hCryptProv = 0;
+
+static int
+win32_urandom_init(int raise)
+{
+    HINSTANCE hAdvAPI32 = NULL;
+    CRYPTACQUIRECONTEXTA pCryptAcquireContext = NULL;
+
+    /* Obtain handle to the DLL containing CryptoAPI. This should not fail. */
+    hAdvAPI32 = GetModuleHandle("advapi32.dll");
+    if(hAdvAPI32 == NULL)
+        goto error;
+
+    /* Obtain pointers to the CryptoAPI functions. This will fail on some early
+       versions of Win95. */
+    pCryptAcquireContext = (CRYPTACQUIRECONTEXTA)GetProcAddress(
+                               hAdvAPI32, "CryptAcquireContextA");
+    if (pCryptAcquireContext == NULL)
+        goto error;
+
+    pCryptGenRandom = (CRYPTGENRANDOM)GetProcAddress(hAdvAPI32,
+                                                     "CryptGenRandom");
+    if (pCryptGenRandom == NULL)
+        goto error;
+
+    /* Acquire context */
+    if (! pCryptAcquireContext(&hCryptProv, NULL, NULL,
+                               PROV_RSA_FULL, CRYPT_VERIFYCONTEXT))
+        goto error;
+
+    return 0;
+
+error:
+    if (raise)
+        PyErr_SetFromWindowsErr(0);
+    else
+        Py_FatalError("Failed to initialize Windows random API (CryptoGen)");
+    return -1;
+}
+
+/* Fill buffer with size pseudo-random bytes generated by the Windows CryptoGen
+   API. Return 0 on success, or -1 on error. */
+static int
+win32_urandom(unsigned char *buffer, Py_ssize_t size, int raise)
+{
+    Py_ssize_t chunk;
+
+    if (hCryptProv == 0)
+    {
+        if (win32_urandom_init(raise) == -1)
+            return -1;
+    }
+
+    while (size > 0)
+    {
+        chunk = size > INT_MAX ? INT_MAX : size;
+        if (!pCryptGenRandom(hCryptProv, chunk, buffer))
+        {
+            /* CryptGenRandom() failed */
+            if (raise)
+                PyErr_SetFromWindowsErr(0);
+            else
+                Py_FatalError("Failed to initialized the randomized hash "
+                        "secret using CryptoGen)");
+            return -1;
+        }
+        buffer += chunk;
+        size -= chunk;
+    }
+    return 0;
+}
+#endif /* MS_WINDOWS */
+
+
+#ifdef __VMS
+/* Use openssl random routine */
+#include <openssl/rand.h>
+static int
+vms_urandom(unsigned char *buffer, Py_ssize_t size, int raise)
+{
+    if (RAND_pseudo_bytes(buffer, size) < 0) {
+        if (raise) {
+            PyErr_Format(PyExc_ValueError,
+                         "RAND_pseudo_bytes");
+        } else {
+            Py_FatalError("Failed to initialize the randomized hash "
+                          "secret using RAND_pseudo_bytes");
+        }
+        return -1;
+    }
+    return 0;
+}
+#endif /* __VMS */
+
+
+#if !defined(MS_WINDOWS) && !defined(__VMS)
+
+/* Read size bytes from /dev/urandom into buffer.
+   Call Py_FatalError() on error. */
+static void
+dev_urandom_noraise(char *buffer, Py_ssize_t size)
+{
+    int fd;
+    Py_ssize_t n;
+
+    assert (0 < size);
+
+    fd = open("/dev/urandom", O_RDONLY);
+    if (fd < 0)
+        Py_FatalError("Failed to open /dev/urandom");
+
+    while (0 < size)
+    {
+        do {
+            n = read(fd, buffer, (size_t)size);
+        } while (n < 0 && errno == EINTR);
+        if (n <= 0)
+        {
+            /* stop on error or if read(size) returned 0 */
+            Py_FatalError("Failed to read bytes from /dev/urandom");
+            break;
+        }
+        buffer += n;
+        size -= (Py_ssize_t)n;
+    }
+    close(fd);
+}
+
+/* Read size bytes from /dev/urandom into buffer.
+   Return 0 on success, raise an exception and return -1 on error. */
+static int
+dev_urandom_python(char *buffer, Py_ssize_t size)
+{
+    int fd;
+    Py_ssize_t n;
+
+    if (size <= 0)
+        return 0;
+
+    Py_BEGIN_ALLOW_THREADS
+    fd = open("/dev/urandom", O_RDONLY);
+    Py_END_ALLOW_THREADS
+    if (fd < 0)
+    {
+        PyErr_SetFromErrnoWithFilename(PyExc_OSError, "/dev/urandom");
+        return -1;
+    }
+
+    Py_BEGIN_ALLOW_THREADS
+    do {
+        do {
+            n = read(fd, buffer, (size_t)size);
+        } while (n < 0 && errno == EINTR);
+        if (n <= 0)
+            break;
+        buffer += n;
+        size -= (Py_ssize_t)n;
+    } while (0 < size);
+    Py_END_ALLOW_THREADS
+
+    if (n <= 0)
+    {
+        /* stop on error or if read(size) returned 0 */
+        if (n < 0)
+            PyErr_SetFromErrno(PyExc_OSError);
+        else
+            PyErr_Format(PyExc_RuntimeError,
+                         "Failed to read %zi bytes from /dev/urandom",
+                         size);
+        close(fd);
+        return -1;
+    }
+    close(fd);
+    return 0;
+}
+#endif /* !defined(MS_WINDOWS) && !defined(__VMS) */
+
+/* Fill buffer with pseudo-random bytes generated by a linear congruent
+   generator (LCG):
+
+       x(n+1) = (x(n) * 214013 + 2531011) % 2^32
+
+   Use bits 23..16 of x(n) to generate a byte. */
+static void
+lcg_urandom(unsigned int x0, unsigned char *buffer, size_t size)
+{
+    size_t index;
+    unsigned int x;
+
+    x = x0;
+    for (index=0; index < size; index++) {
+        x *= 214013;
+        x += 2531011;
+        /* modulo 2 ^ (8 * sizeof(int)) */
+        buffer[index] = (x >> 16) & 0xff;
+    }
+}
+
+/* Fill buffer with size pseudo-random bytes, not suitable for cryptographic
+   use, from the operating random number generator (RNG).
+
+   Return 0 on success, raise an exception and return -1 on error. */
+int
+_PyOS_URandom(void *buffer, Py_ssize_t size)
+{
+    if (size < 0) {
+        PyErr_Format(PyExc_ValueError,
+                     "negative argument not allowed");
+        return -1;
+    }
+    if (size == 0)
+        return 0;
+
+#ifdef MS_WINDOWS
+    return win32_urandom((unsigned char *)buffer, size, 1);
+#else
+# ifdef __VMS
+    return vms_urandom((unsigned char *)buffer, size, 1);
+# else
+    return dev_urandom_python((char*)buffer, size);
+# endif
+#endif
+}
+
+void
+_PyRandom_Init(void)
+{
+    char *env;
+    void *secret = &_Py_HashSecret;
+    Py_ssize_t secret_size = sizeof(_Py_HashSecret);
+
+    if (random_initialized)
+        return;
+    random_initialized = 1;
+
+    /*
+      By default, hash randomization is disabled, and only
+      enabled if PYTHONHASHSEED is set to non-empty or if
+      "-R" is provided at the command line:
+    */
+    if (!Py_HashRandomizationFlag) {
+        /* Disable the randomized hash: */
+        memset(secret, 0, secret_size);
+        return;
+    }
+
+    /*
+      Hash randomization is enabled.  Generate a per-process secret,
+      using PYTHONHASHSEED if provided.
+    */
+
+    env = Py_GETENV("PYTHONHASHSEED");
+    if (env && *env != '\0' & strcmp(env, "random") != 0) {
+        char *endptr = env;
+        unsigned long seed;
+        seed = strtoul(env, &endptr, 10);
+        if (*endptr != '\0'
+            || seed > 4294967295UL
+            || (errno == ERANGE && seed == ULONG_MAX))
+        {
+            Py_FatalError("PYTHONHASHSEED must be \"random\" or an integer "
+                          "in range [0; 4294967295]");
+        }
+        if (seed == 0) {
+            /* disable the randomized hash */
+            memset(secret, 0, secret_size);
+        }
+        else {
+            lcg_urandom(seed, (unsigned char*)secret, secret_size);
+        }
+    }
+    else {
+#ifdef MS_WINDOWS
+        (void)win32_urandom((unsigned char *)secret, secret_size, 0);
+#else /* #ifdef MS_WINDOWS */
+# ifdef __VMS
+        vms_urandom((unsigned char *)secret, secret_size, 0);
+# else
+        dev_urandom_noraise((char*)secret, secret_size);
+# endif
+#endif
+    }
+}

Python/sysmodule.c

@@ -1126,6 +1126,7 @@ static PyStructSequence_Field flags_fields[] = {
     /* {"unbuffered",                   "-u"}, */
     /* {"skip_first",                   "-x"}, */
     {"bytes_warning", "-b"},
+    {"hash_randomization", "-R"},
     {0}
 };
 
@@ -1134,9 +1135,9 @@ static PyStructSequence_Desc flags_desc = {
     flags__doc__,       /* doc */
     flags_fields,       /* fields */
 #ifdef RISCOS
-    12
+    13
 #else
-    11
+    12
 #endif
 };
 
@@ -1169,6 +1170,7 @@ make_flags(void)
     /* SetFlag(saw_unbuffered_flag); */
     /* SetFlag(skipfirstline); */
     SetFlag(Py_BytesWarningFlag);
+    SetFlag(Py_HashRandomizationFlag);
 #undef SetFlag
 
     if (PyErr_Occurred()) {