Getting Pyrit to work on my system and other WPA-PSK related rand

Recently i started reading about WPA-PSK security, as a client was using it, in their office, and after some discussion, wanted to see how (in)secure might be.

After some reading it looks like the best way to hack a WPA-PSK based system (for the moment) is to create a rainbow table of possible PMK's (Pairwise Master Key) and then let loose tools like cowpatty and aircrack-ng.
Basically you exchange time for space, as the same PMK can be used on any AP with same SSID.

The problem is, that a PMK requires something in the range of 16.000+ rounds of SHA-1 and this requires some really big computing power.
To give you an idea of the computing power required, my quad core@3Ghz can do about 1.2K PMK's/sec,which is not that bad, but will take weeks to go over a descent word list.

Doing some more search i found that there is a program called Pyrit, that uses the power of the GPU to do some serious PMK crunching. My "vanilla" (not over clocked) Nvidia 8800 GT does 4.800/sec while other, newer Nvidia cards can reach close to 50.000/sec PMK's

UPDATE 17/5/2009
The problem mentioned below is solved in revision 99 of Pyrit
Looks like it was a CUDA 2.2 bug
I am leaving the text for 'historical reasons' but you can safely ignore the fix

So i updated my CUDA drivers and SDK (Pyrit requires CUDA 2,2), got Pyrit from the SVN, build it and run my first benchmark, using only the CPU's.
Things were good, so i moved on to build the Nvidia CUDA module for Pyrit.
The build was ok

stelios@Athena:~/pyrit/pyrit-read-only/cpyrit_cuda$ ./setup.py build
running build
running build_ext
Compiling CUDA module using nvcc 2.2, V0.2.1221...
ptxas info : Compiling entry function 'cuda_pmk_kernel'
ptxas info : Used 42 registers, 32+24 bytes smem, 12 bytes cmem[1]
Building modules...
stelios@Athena:~/pyrit/pyrit-read-only/cpyrit_cuda$ sudo ./setup.py
install
running install
running build
running build_ext
Skipping rebuild of Nvidia CUDA kernel ...
Building modules...
running install_lib
running install_egg_info
Removing /usr/lib/python2.5/site-packages/CPyrit_CUDA-0.2.3.egg-info
Writing /usr/lib/python2.5/site-packages/CPyrit_CUDA-0.2.3.egg-info

but when trying to run the benchmark again i got an error

stelios@Athena:~/pyrit/pyrit-read-only/cpyrit_cuda$ pyrit benchmark
Pyrit 0.2.3 (C) 2008, 2009 Lukas Lueg http://pyrit.googlecode.com
This code is distributed under the GNU General Public License v3

The ESSID-blobspace seems to be empty; you should create an ESSID...

Failed to load CUDA-core (CUDA_ERROR_INVALID_IMAGE).
Running benchmark for at least 60 seconds...

CPU-Core (x86_64): 302.43 PMKs/s, 99.41% occupancy
CPU-Core (x86_64): 292.03 PMKs/s, 90.08% occupancy
CPU-Core (x86_64): 300.92 PMKs/s, 87.42% occupancy
CPU-Core (x86_64): 303.17 PMKs/s, 99.17% occupancy

Benchmark done. 1198.55 PMKs/s total.

For some reason the CUDA part was failing to load.

Googling about the error, i found a couple others had the same issue, so it was not just me, doing something wrong.
I emailed the author, but received no reply, so after a day started looking at the code to see where the problem comes from. Open Source rulez :)

It turned out that the module failed to load the CUDA kernel.
Pyrit "converts" the CUDA cubit module to an include file _cpyrit_cudakernel.cubin.h and then uses the CUDA API to load the kernel module.
In my case ,For some reason the _cpyrit_cudakernel.cubin.h seems to have an invalid
cuda kernel image.

So i changed the part that loads the kernel in _cpyrit_cuda.c from the include file

ret = cuModuleLoadData(&self->mod, &__cudakernel_module);

to

ret = cuModuleLoad(&self->mod, "/your/path/to/cubitfile/_cpyrit_cudakernel.cubin");

(P.S add the correct path to your cubit file)

that loads the cubit file directly.

That got the problem fixed and benchmark worked like a charm

stelios@Athena:~/pyrit/pyrit-read-only/cpyrit_cuda$ pyrit benchmark
Pyrit 0.2.2 (C) 2008, 2009 Lukas Lueg http://pyrit.googlecode.com
This code is distributed under the GNU General Public License v3

Running benchmark for at least 60 seconds...

CUDA-Device #1 'GeForce 8800 GT': 4796.11 PMKs/s, 89.75% occupancy
CPU-Core (x86_64): 283.45 PMKs/s, 84.37% occupancy
CPU-Core (x86_64): 298.66 PMKs/s, 96.09% occupancy
CPU-Core (x86_64): 289.44 PMKs/s, 99.15% occupancy

Benchmark done. 5667.66 PMKs/s total.

Haven't looked at why the _cpyrit_cudakernel.cubin.h has a corrupted
kernel, will probably do so, during the weekend and post any patches to fix it.