john-users - OpenCL build segfaults

[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Date: Sun, 31 Jul 2011 17:40:46 +0200
From: groszek <necro@...icon.pl>
To: john-users@...ts.openwall.com
Subject: OpenCL build segfaults

Hi.
I've built the older 1.7.7-jumbo-1, patched with "OpenCL support for
cracking NT, raw-MD4, raw-MD5, NSLDAP and raw-SHA1 hashes".

When I do ./john --test --format=raw-md5-opencl it works:
  Benchmarking: Raw MD5 [raw-md5-opencl]...
  Kernel path is : /usr/share/john/md5_opencl_kernel.cl

  OpenCL Platform: <<<NVIDIA CUDA>>> and device: <<<GeForce GTS 250M>>>
  DONE
  Raw:    24672K c/s real, 24672K c/s virtual

Same with ./john --test --format=raw-sha1-opencl:
  Benchmarking: Raw SHA-1 OpenCL [SHA-1]...
  Kernel path is : /usr/share/john/sha1_opencl_kernel.cl

  OpenCL Platform: <<<NVIDIA CUDA>>> and device: <<<GeForce GTS 250M>>>
  DONE
  Many salts:     6618K c/s real, 6618K c/s virtual
  Only one salt:  7137K c/s real, 7137K c/s virtual

However, when I actually want to crack something with them it segfaults:
  Kernel path is : /usr/share/john/md5_opencl_kernel.cl

  OpenCL Platform: <<<NVIDIA CUDA>>> and device: <<<GeForce GTS 250M>>>
  Loaded 3439 password hashes with no different salts (Raw MD5
[raw-md5-opencl])
  Remaining 3437 password hashes with no different salts
  Segmentation fault

And same goes for sha1. Just before it segfaults it hangs for few
seconds, with 100% cpu usage.

I have x86_64 build, compiled with gcc x86_64-pc-linux-gnu-4.5.2, and
CUDA toolkit/sdk version 4.0.8.

I also tried to integrate this patch into newest 1.7.8-jumbo-4 but
failed miserably, to a point where self-checks of those algorithms
didn't pass. Maybe I'll give it another go later. Are there any plans on
getting OpenCL support beyond basic 3rd party patch?

Any idea, anyone experienced that? Thanks.