phc-discussions - Re: [PHC] Quick gripe... in case there's ever another contest

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20140405191448.GA8325@bolet.org>
Date: Sat, 5 Apr 2014 21:14:48 +0200
From: Thomas Pornin <pornin@...et.org>
To: Poul-Henning Kamp <phk@....freebsd.dk>
Cc: discussions@...sword-hashing.net
Subject: Re: [PHC] Quick gripe... in case there's ever another contest

On Sat, Apr 05, 2014 at 06:52:28PM +0000, Poul-Henning Kamp wrote:
> Password scrambling is a systems function, systems are written in C.

However, I am not talking about the _deployed_ implementation, which
will be written in whatever is most convenient and efficient on a
specific architecture. I am talking about the reference implementation,
which is not meant to actually hash user passwords, but to serve as the
testbed for the deployed implementation. The reference implementation
should work everywhere, always producing the same outputs on the same
inputs.

If you prefer, the _reference_ implementation is the "portable" one,
which works everywhere but (possibly) slowly, while the _deployed_
implementation is the "optimized" one, which runs fast on a specific
architecture but (possibly) incorrectly on others. When you have
actual passwords to hash, you want to use the optimized code, not the
portable code, because password hashing is a muscle contest between
attacker and defender; as the defender, you want muscle.

In fact, the reference implementation should not run anywhere except on
the implementer's desktop system, to help him produce the optimized
implementation. Similarly, the reference implementation shall be
readable, clear, even if clarity comes at the expense of performance.

Of course, if the optimized implementation is ALSO portable and
readable, so much the better, but that's not its main role.

My point is that you will rarely get such a reference implementation
(works everywhere) with C. It is possible to write portable C code, but
that's quite beyond simply sticking to "plain ISO C"; you also have to
avoid all the undefined and implementation-defined behaviours that C is
so full of. E.g. don't do unaligned accesses or anything which violates
type aliasing rules (so any endian-sensitive trick is right out); don't
right-shift a negative integer; don't actually overflow or underflow a
signed computation; stick to unsigned types with definite sizes (such as
uint32_t) or add masks everywhere; and so on. This can be done.

But it is utopic to hope that the average "reference code" submitted by
an algorithm designer will achieve such portability. In practice, if you
take a random submitter's code from a crypto competition, you can
already consider yourself lucky if it works on BOTH a 32-bit and a
64-bit PC running Linux + gcc and "basic" optimization flags. If it also
works on a Sparc system (big-endian, does not tolerate unaligned
accesses), open some Champagne to celebrate.

(Which all amounts to say that one-size-fits-all libraries, portable AND
readable AND optimized, are a rarity.)

	--Thomas Pornin