Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 15 Jun 2019 19:39:56 +0200
From: Solar Designer <solar@...nwall.com>
To: oss-security@...ts.openwall.com
Cc: security@...tpractical.com
Subject: Re: Apache::Session's use of md5 and more

Hi,

On Sat, Jun 15, 2019 at 05:09:53PM +0200, Raphael Geissert wrote:
> I just stumbled upon Apache::Session's Generate::MD5 module, which
> appears to be used to generate the session ids for cookies and the
> like.
> 
> Not only does it use MD5,

Which is perfectly fine for this use case, except that it distracts
attention from real issues, so might need to be "fixed" to be e.g.
SHA-256 for that reason.

Let's not confuse technical and psychological aspects.

> but its source of entropy is weak

That's the real issue.

> and does two rounds of hashing.

This is fine, but can be optimized out along with the move to SHA-256.

> From the source code[1]:
> 
>     $session->{data}->{_session_id} =
>         substr(Digest::MD5::md5_hex(Digest::MD5::md5_hex(time(). {}.
> rand(). $$)), 0, $length);
> 
> (where $length is 32 by default)

This uses 3 or 4 pieces of data: time in seconds since Unix epoch, an
address within the process, whatever seed rand() was initialized with
(might also be time, or not), and PID.  This might be insufficient to
prevent successfully inferring these inputs from the hash (by probing
likely inputs), in which case this also leaks these inputs - kind of a
remote ASLR leak, which matters if the process is persistent, etc. - on
top of the more obvious impact of being able to predict session IDs.

Also, does this generate unique session IDs if called twice in a row
from the same process?  It appears that due to the "{}" and the "rand()"
call it usually does, but perhaps not reliably to an extent where we'd
rely on that for security.

> Am I missing something, or has this code actually been in use for ages
> and gone unnoticed ? I couldn't find any CVE for this.
> 
> So far I found this reference, but only mentions the use of MD5 as a weakness:
> https://gitlab.ow2.org/lemonldap-ng/lemonldap-ng/issues/695

That thread focuses on MD5 to an extent where everyone in there seems to
think that replacing MD5 with SHA-256 would magically fix whatever issue
they're thinking there is.  They're wrong.

This is especially surprising given that Nuel Guillaume who opened the
issue writes in one of the comments (5 years ago):

| We can easily determine time()
| 
| {} A memory allocation for a hash (dict).
| The output looks like "HASH (0x97b27ec)."
| The last 3 characters: "7EC" are fixed to each machine.
| 
| Rand Perl function calls directly to the rand () function in libc.
| rand() is not secure at all. Just find 30 values of rand() to determine the srand (the seed).
| But it can be easier if we have Perl prior to 5.004. ( => srand(time() ).
| If we have Perl 5.004 or upper, /dev/urandom is used for the default srand.
| See:
| http://turtle.ee.ncku.edu.tw/docs/perl/manual/pod/perlfunc/srand.html
| http://stackoverflow.com/questions/12497045/what-are-the-weaknesses-of-perls-srand-default-seed-post-version-5-004
| 
| $$ Is the PID of the process and his value is between 1000 and 32768 (/proc/sys/kernel/pid_max = 32768)

but then even with this understanding goes on to suggest merely "Replace
md5 with SHA1 or SHA256 and keep your Perl version update."

I didn't review Perl's rand(), but apparently Nuel thought the
initialization from /dev/urandom on newer Perl somehow made rand() safe
from having its seed inferred?  I doubt this is the case, as I expect
the seed and/or the internal state is tiny either way.  And I doubt it
takes as many as "30 values of rand() to determine the srand (the
seed)."  I'd expect 1 to be enough.  But we need to review the code
before making any claims.

...OK, I just took a look.  Perl's util.c: Perl_seed() reads just 32
bits from /dev/urandom, with compile-time and runtime fallbacks to
gettimeofday() and getpid() and some more ASLR leaks.  (Fun fact: the
fallbacks will also occur when the 32-bit value read from /dev/urandom
just happens to be 0.  As a result, the seed is almost never a 0.)

> From a quick look at the reverse dependencies of the Debian package,
> there are some users of Apache::Session:
> * RequestTracker (RT) : from a quick look at the session id in the
> cookie set by rt.cpan.org I'd say it does use Generate::MD5
> * Torrus: no idea if the Generate::MD5 module is used
> * LemonLdap::NG : they replaced Generate::MD5 by a similar code using
> SHA256, but still using two rounds of hashing
> 
> CC'ing BestPractical. Will open an issue on LemonLdap::NG's gitlab.

Please focus on lack of (explicit) use of a CSPRNG such as /dev/urandom,
not on use of MD5 nor the double-hashing (which are non-issues).

> [1]https://metacpan.org/source/CHORNY/Apache-Session-1.93/lib/Apache/Session/Generate/MD5.pm

Alexander

Powered by blists - more mailing lists

Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.