Date: Fri, 9 Aug 2013 16:10:42 +0200 From: Katja Malvoni <kmalvoni@...il.com> To: john-dev@...ts.openwall.com Subject: Re: Parallella: bcrypt On Fri, Aug 9, 2013 at 3:38 PM, Solar Designer <solar@...nwall.com> wrote: > Katja, > > On Mon, Aug 05, 2013 at 07:19:07PM +0200, Katja Malvoni wrote: > > I run self test until it stalled. It happens in while loop on host side. > > One of the cores (first time it was 14, second time 13, third time 15) > > stalls somewhere before do while loop. I checked shared DRAM, 16 is > written > > in memory for both start arrays for that specific core so conditions in > > while loops are true but it never reaches BF_crypt() call nor resetting > > start flags. I also added writing different number instead of result > after > > every line in main on Epiphany side - core that stalls never reached > first > > write that was before do while loop. Only zeroes were in memory for > result > > and core_done flag for that core. All the other cores except the one that > > stalls work correct. And it happens only in first cryp_all call. So far > it > > never happened in the middle of computation. It seems as that one core > > never starts computation. > > This is a separate problem from the other reliability problem you had > reported (sometimes cracking one password less than expected), right? > So we've got two reliability problems in the current code, right? > > What do you intend to do about these? > Yes, there are two problems. For stalling problem - I'll run it until it stalls and try to figure out where exactly does it happen. When I identify where it stalls I think I'll know is it because reset fails or it's something else. Also, I'll try resetting cores one by one using e_reset_core() instead of e_reset_system(). On Fri, Aug 9, 2013 at 3:41 PM, Solar Designer <solar@...nwall.com> wrote: > I think it'll be easier for you to debug this if you create a program > that will verify each and every computed bcrypt hash. (As discussed > before, this is not what happens when we're cracking passwords. > Although we do use every bcrypt computation results, most failures can > go undetected.) > > With a 100%-verifying program, you should be able to trigger the issue > much more quickly and more reliably, so you'd be able to test different > theories as to its cause quicker too. > OK, I'll do that. Katja Content of type "text/html" skipped
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.