Re: [RFC][PATCH] Pseudo-random number generator

[Jan Glauber <jan.glauber@de.ibm.com>]

On Mon, 2006-12-04 at 11:15 -0500, Valdis.Kletnieks@vt.edu wrote:
> On Fri, 01 Dec 2006 14:19:15 +0100, Jan Glauber said:
> > New s390 machines have hardware support for the generation of pseudo-random
> > numbers. This patch implements a simple char driver that exports this numbers
> > to user-space. Other possible implementations would have been:
> 
> > +	for (i = 0; i < 16; i++) {
> > +		entropy[0] = get_clock();
> > +		entropy[1] = get_clock();
> > +		entropy[2] = get_clock();
> > +		entropy[3] = get_clock();
> 
> By the time this loop completes, we'll have done 64 get_clock() - and if an
> attacker has a good estimate of what the system clock has in it, they'll be
> able to guess all 64 values, since each pass through the loop will have fairly
> predictable timing.  So as a result, the pseudo-random stream will be a *lot*
> less random than one would hope for...

I completely agree. Filling the input buffer with timestamps looks quite
uncomfortable but was exactly what the hardware specification suggested.

At least for the initialisation of the PRNG I preferred get_random_bytes()
over get_clock to get a good initial seed. But get_random_bytes cannot
be used during normal operation since the PRNG read should not block.

> > +		/*
> > +		 * It shouldn't weaken the quality of the random numbers
> > +		 * passing the full 16 bytes from STCKE to the generator.
> > +		 */
> 
> As long as you realize that probably 12 or 13 or even more of those 16 bytes
> are likely predictable (depending exactly how fast the hardware clock ticks),
> and as a result the output stream will also be predictable.

Yes, if an attacker knows the initial clock value a brute-force attack
would be feasible to predict the output. But I don't know if the
hardware completely relies on the clock values or if there is any
internal state which is not visible by an attacker. I will try to find
out more details...

Jan