[PATCH] crypt_blowfish 1.2

• • For actual implementation, we set an array index in the variable "bug"
• • (0 means no bug, 1 means sign extension bug emulation) and a flag in the
• • variable "safety" (bit 16 is set when the safety measure is requested).
• • Valid combinations of settings are:
• • Prefix "$2a$": bug = 0, safety = 0x10000
• • Prefix "$2x$": bug = 1, safety = 0
• • Prefix "$2y$": bug = 0, safety = 0

If I'm understanding the change correctly, we should now be advising
users to transition their code to '$2y$' rather than '$2a$', with

Yes, but this is not terribly important. In practice, "$2a$" is almost
the same as "$2y$". For passwords that don't contain the '\xff'
character (which is not even valid in UTF-8 sequences), these two are
100% equivalent. For realistic passwords that do contain this
character, I had one "hit" in 150,000+ such passwords tested:

http://www.openwall.com/lists/oss-security/2011/07/08/1

So this is negligible, and even for the affected passwords (where "$2y$"
and "$2a$" hashes differ by more than just the prefix) this only matters
if those password hashes are ever migrated to other systems (non-PHP).

The reason why I went for this is that I consider the security advantage
of avoiding easy collisions with the buggy hashes non-negligible.

perhaps a note mentioning the '$2x$' prefix for "transitioning users
with passwords that contain non-ASCII characters with the 8th bit set".

We need to be careful here such that no one starts using this for newly
set passwords. This bit of documentation should be available to those
few who actually need it (I expect that most sites won't care), but
maybe it should not be on the function crypt() documentation page.

Obviously, any documentation change in this regard will need to be
pending on the version these patches get rolled into...

Yes - need to release PHP versions with this code first.

Thanks,

Alexander

Hi,

Thanks for the patches, applied to all active branches.

Thank you!

About the tests, it would be very good to have them ported as phpt. As
far as I remember I did that back then when I first ported it to php.

No, your tests appear to be specific to PHP's peculiarities. The tests
ported from crypt_blowfish's wrapper.c only appear in 5.4, from
crypt_blowfish 1.1 - probably Stas ported them (thanks!)

Anyway, attached are patches adding the tests from crypt_blowfish 1.2 to
trunk, 5.4, and 5.3. These patches also fix a bug in crypt.c.

I did not port tests of behavior on error, though. Some of these are
redundant with other tests that you already have in other .phpt files,
others would not fit in the existing array() format - would require a
separate loop, which I was too lazy to add. The tests that I did not
port are these (from my wrapper.c):

{"*0", "", "$2a$03$CCCCCCCCCCCCCCCCCCCCC."},
{"*0", "", "$2a$32$CCCCCCCCCCCCCCCCCCCCC."},
{"*0", "", "$2z$05$CCCCCCCCCCCCCCCCCCCCC."},
{"*0", "", "$2`$05$CCCCCCCCCCCCCCCCCCCCC."},
{"*0", "", "$2{$05$CCCCCCCCCCCCCCCCCCCCC."},
{"*1", "", "*0"},

Alexander