Random number generator in Python core

Python/hash.c

+#include "Python.h"
+#include "pyrandom.h"
+
+Py_hash_t Py_RndHashSeed = -1;

[gregory.p.smith, 2012/01/15 02:18:38]

This should be a Py_uhash_t and the default value should be 0 so that the
behavior if it is never set to anything else is the same as today's non-seeded
behavior.

Instead of being a global, should this be a member of PyInterpreterState and be
initialized from pystate.c's PyInterpreterState_New()?  For 3.3 I think that
makes sense.  A process with multiple embedded Python interpreters would have
different hash seeds for each of them.

For backports it will have to be a global to avoid changing the
PyInterpreterState definition (API).

+
+void
+_Py_InitRndHashSeed(void)
+{
+    static int initialized = 0;
+    int result;
+    if (initialized) {
+        return;
+    }
+    initialized = 1;
+
+    result = PyOS_URandom((unsigned char*)&Py_RndHashSeed, sizeof(Py_hash_t));
+    /* fall back to Mersenne twister */
+    if (result == -1) {
+        _Py_MT_RandomState state;
+        time_t seed;
+        int i;
+
+        time(&seed);
+        _Py_MT_GenRand_Init(&state, seed);

[gregory.p.smith, 2012/01/15 02:18:38]

Include the pid, system uptime and any other easy to get environmental info into
the seed.  No sane OS is going to use this code path...

But it should be better than simply time if at all possible as the time an
interpreter serving something on a network started up is often easy to
determine.

+
+        Py_RndHashSeed = _Py_MT_GenRand_Int32(&state);
+        if (Py_RndHashSeed == -1) {
+            Py_RndHashSeed = -2;
+        }
+
+    }
+}
+
+/* Randomized hash seed */
+Py_hash_t
+Py_GetRndHashSeed(void)
+{
+    return Py_RndHashSeed;
+}