phc-discussions - (not) protecting password length from side-channels (Re: [PHC] Argon2 modulo division)

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: <20150423004705.GC31523@openwall.com>
Date: Thu, 23 Apr 2015 03:47:05 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: (not) protecting password length from side-channels (Re: [PHC] Argon2 modulo division)

On Wed, Apr 22, 2015 at 09:13:21AM -0700, Bill Cox wrote:
> It turns out that there was not a single entry in the competition that is
> power rail analysis resistant

... with respect to password length.

I'm with Thomas on this.  It is futile for PHC candidates to fully
protect the password length.  (They should, however, make their running
time _mostly_ independent of password length, and Milan's benchmarks
confirm that they do.)  Yet the callers may, even with the current
PHS() API, such as by padding the password as Thomas suggested (and
preferably also encoding the original password length in there), if they
are somehow able to perform this padding in a side-channel safe way
(e.g., before reaching the device where side-channels matter).

Even if the length were not almost always already slightly-leaked in the
caller, just how would a PHC candidate protect it fully when called with
a pointer to string and a length?  With something similar to bcrypt's
key setup, repeating the password until a buffer is full?  That is still
leaky, even if possibly slightly less so (or not, because the length
signal is amplified through affecting the read pointer multiple times)
than e.g. a memcpy() is.  The memory access pattern still varies by
password length, even if the number of characters read (including
duplicate reads) is constant.  Permitting PHS() to read from the
password buffer beyond the provided length would be quite a different
API, and a rather dangerous one (many callers would not care to provide
room for the over-read).  It is better to let the paranoid callers pad
the password input with the current API than to subject everyone to a
more dangerous API that would only benefit a minority.

A similar issue arises once in a while in discussions around constant
time string comparison features being added to libraries and languages.

Alexander