Three format fails after merging GPU mask branch

[magnumripper]

Testing: bfegg, Eggdrop [Blowfish 32/64]... (4xOMP) FAILED (get_key(1))
Testing: HMAC-SHA1 [password is key, SHA1 128/128 XOP 8x]... (4xOMP) FAILED (get_key(3))
Testing: mysql, MySQL pre-4.1 [32/64]... (4xOMP) FAILED (get_key(4))

This may or may not be problems in the formats. It's more likely these are "self test technical" issues that need to be worked around, either in formats.c or in these formats.

[sayan1an]

Yes, I saw that too. Looking into it right now.

[sayan1an]

Quick fix:
change line 539 in format.c to the following:

for (i = match - 1; i >= 0; i--)

[sayan1an]

However, it looks like the these format generate same hash for multiple passwords. Otherwise, there shouldn't have been any problem.

[magnumripper]

Hmm or maybe they just have dupe test vectors?

[magnumripper]

Hm, no.

$ ../run/john -form:bfegg --list=format-tests 
bfegg   0   +9F93o1OxwgK1   123456
bfegg   1   +C/.8o.Wuph9.   qwerty
bfegg   2   +EEHgy/MBLDd0   walkman
bfegg   3   +vPBrs07OTXE/   tesztuser
bfegg   4   +zIvO/1nDsd9.   654321

[magnumripper]

Are we always setting all keys in the batch? If not, you might see leftovers from a previous crypt.

[sayan1an]

Yes, right. But then, shouldn't the format cleanse its buffer such that cmp_one() and cmp_exact() does not return false positives?

Anyway, the latest commit emulates the self-test exactly as we did before for older formats and externally, they shouldn't notice any difference.

[magnumripper]

shouldn't the format cleanse its buffer such that cmp_one() and cmp_exact() does not return false positives?

In real runs, we always set n keys and then call crypt_all(n) and cmp_*(n) so buffer cleaning would just harm performance. We should try to make the self-test work as close to real-life as possible (perhaps fill keys with dummy strings that are not included in any actual test vector).