Keepass OpenCL format.

[magnumripper]

No description provided.

[magnumripper]

Also fixes a bug in shared valid() - it didn't reject unsupported hashes :-(

@kholia this is important. We really really need to reject unsupported hashes, always and forever. In this case the [CPU] format would bail out in an awkward way - which is better than nothing - but the correct behavior is rejecting them in valid().

[kholia]

Thanks @magnumripper. I missed this opportunity to make valid correct and safe.

[magnumripper]

I really can't see WTH is wrong with this format. On a few devices, it works like a champ. On others it fails. On some it works fine with LWS=1 (this might be a clue) but not otherwise.

I have stared at the code for 24h to no avail. Something must be amiss, but WHAT!? @kholia can you try it out and see if whatever GPU and driver you use work or not, or if it happens to reveal some clue?

[magnumripper]

This is interesting. Fails a normal test:

$ ../run/john -form:keepass-opencl -test=0 -dev=1 
Device 1: HD Graphics 4000
Testing: KeePass-opencl [SHA256 AES/Twofish/Chacha OpenCL]... FAILED (cmp_all(2))

...but passes with LWS=1:

$ LWS=1 ../run/john -form:keepass-opencl -test=0 -dev=1 
Device 1: HD Graphics 4000
Testing: KeePass-opencl [SHA256 AES/Twofish/Chacha OpenCL]... PASS

...however when using --test-full instead, it works fine with any LWS:

$ ../run/john -form:keepass-opencl -test-full=0 -dev=1 
Device 1: HD Graphics 4000
Testing: KeePass-opencl [SHA256 AES/Twofish/Chacha OpenCL]... PASS

[kholia]

Quick notes,

$ ../run/john --format=keepass-opencl --test
Device 0: GeForce GTX 1050 Ti
Benchmarking: KeePass-opencl [SHA256 AES/Twofish/Chacha OpenCL]... FAILED (cmp_all(1))

$ ../run/john --format=keepass-opencl -test-full=0
Device 0: GeForce GTX 1050 Ti
Testing: KeePass-opencl [SHA256 AES/Twofish/Chacha OpenCL]... FAILED (cmp_all(49))

$ dpkg -l | grep -i nvidia
...
ii  nvidia-375 ... 375.66-0ubuntu0.17.04.1 ...

Update:

Disabling Twofish and ChaCha test vectors doesn't seem to improve the situation. I will focus on getting salt->version == 1 and salt->algorithm == 0 combination to work first.

[magnumripper]

It works perfectly fine with nvidia drivers 384.90 and 384.111 (Ubuntu packages)

It also works fine with Intel’s CPU driver 1.2.0.42.

I think we’re seeing driver bugs.

[magnumripper]

I tried changing to a whole different AES implementation (not pushed here) but it didn't change a thing with the failing runtimes.

[magnumripper]

@kholia I found that the other AES code we have (introduced with WPA CMAC) is 10x faster. It doesn't seem to fix the issues but anyway I will re-work the whole shebang and issue a new PR later.

[kholia]

All the best for the next attempt @magnumripper! I was hoping to trace the OpenCL kernel to see where things are going wrong but I ran out of debugging energy.