john-dev - Re: Parallella: bcrypt

Follow @Openwall on Twitter for new release announcements and other news

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

Message-ID: <20130520165752.GA10930@openwall.com>
Date: Mon, 20 May 2013 20:57:52 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: Parallella: bcrypt

Hi Katja,

On Mon, May 20, 2013 at 06:45:26PM +0200, Katja Malvoni wrote:
> I tried to reuse the existing Openwall bcrypt implementation since
> parallella supports C. I first tried the simplest approach, load whole
> bcrypt implementation onto the core and execute. The problem is:
> core_test.elf section `.text' will not fit in region `INTERNAL_RAM'
> region `INTERNAL_RAM' overflowed by 6528 bytes

Note that you do not have to put the entire bcrypt onto Epiphany.  At
least initially, it is sufficient to put the configurable-cost loop from
bcrypt onto Epiphany, leaving the rest on host (yes, this means 4 KB
data transfers back and forth, but that's OK for now).  If you're
looking at crypt_blowfish-1.2/crypt_blowfish.c, that loop would be:

	do {
		int done;
[...]
			data.ctx.P[17] ^= tmp2;
		} while (1);
	} while (--count);

Note that this loop contains only one instance of BF_body, which in
turn contains two instances of BF_ENCRYPT (unrolled Blowfish).  In
BF_crypt(), there are a total of 6 instances of BF_ENCRYPT, including
the two inside BF_body in the loop above.  Thus, this change
(implementing just the main loop instead of the entire bcrypt on
Epiphany) should reduce code size to put on Epiphany by a factor of 3.

Please give this a try.

As to other issues you mentioned, Yaniv is the one to comment on them.

Thanks,

Alexander

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.