Re: fe-secure.c and SSL/TLS - Mailing list pgsql-hackers

Hi Marko,

Forgive me for cherry picking two of these...

> I think Postgres uses SSL_VERIFY_PEER + SSL_set_verify() callback instead.
> At least for me, the psql -d "dbname=foo sslmode=verify-ca" fails
> when cert does not match.
I can't comment on the use of psql. My apologies for my ignorance.
However, here's what I see in fe-secure.c around line 695:

static int
verify_cb(int ok, X509_STORE_CTX *ctx)
{   return ok;
}

If `ok` is 0, then validation fails. To learn of the failure, a
program must call SSL_get_verify_result to fetch the error code. Error
codes are listed at https://www.openssl.org/docs/apps/verify.html.

If `ok` is always 1, then validation succeeds. To learn of the
success, a program must call SSL_get_verify_result and ensure
X509_V_OK is returned.

I know of no other ways to check the result of OpenSSL's chain
validation. The open question (for me) is where are
SSL_get_verify_result/X509_V_OK checked? Neither show up in the
Postgres sources.

> 1) OpenSSL guys change default back to TLSv1.0+.
> 2) OpenSSL guys give switch to make TLSv1_method() mean TLSv1.0+.
Well, I don't think that's going to happen, but I could be wrong :)

For what its worth, I agree with you. I want a TLSv1.0+ option and
even had this discussion with Tim Hudson offline.

> 3) libpq starts using TLSv1_2_method() by default.
> 4) libpq will give switch to users to request TLSv1.2.
This might have negative effects on non-TLSv1.2 clients. For example,
an Android 2.3 device can only do TLSv1.0 (IIRC). I think there's a
similar limitation on a lot of Windows XP clients (depending on the IE
version and SChannel version). And OpenSSL-based clients prior to
1.0.0h (released 14 Mar 2012) will have trouble (if I am reading the
change log correctly).

I believe the "standard" way of achieving TLS1.0 and above is to use
the SSLv23_client_method() and then remove the SSL protocols with
SSL_OP_NO_SSLv2 and SSL_OP_NO_SSLv3. I have to use handwaiving around
"standard" because I don't believe its documented anywhere (one of the
devs told me its the standard way to do it.).

Jeff

On Fri, Nov 29, 2013 at 3:19 PM, Marko Kreen <markokr@gmail.com> wrote:
> Reply to mails in pgsql-bugs:
>
>   http://www.postgresql.org/message-id/CAH8yC8mc_2J2UY0Q42WQdWFyaoqT3onG+83Fr=vN46J5+ML94g@mail.gmail.com
>
> and
>
>   http://www.postgresql.org/message-id/CAH8yC8nZVUyCQznkQd8=ELMM4k_=uXJRjt8YF9V22Cy2x_dDjQ@mail.gmail.com
>
>
> * Default ciphersuite
>
>> I would argue nothing should be left to chance, and the project should
>> take control of everything. But I don't really have a dog in the fight ;)
>
> Indeed, on my own servers I've stopped bothering with pattern-based
> ciphersuite strings, now I am listing ciphers explicitly.
>
> But the discussion here is about default suite for admins who don't
> know or care about TLS.  Also, it would be good if it does not
> need to be tuned yearly to stay good.
>
>
> * SSL_get_verify_result
>
> I think Postgres uses SSL_VERIFY_PEER + SSL_set_verify() callback instead.
> At least for me, the psql -d "dbname=foo sslmode=verify-ca" fails
> when cert does not match.
>
>
> * SNI (Server Name Indication extension).
>
> It might be proper to configure it for connections, but it's unlikely
> to be useful as the many-domains-few-ips-one-port problem really does
> not apply to databases.  And from my experience, even the non-SNI
> hostname check is underused (or even - unusable) in many database
> setups.
>
>
> * TLSv1.2
>
> That's the remaining problem with Postgres SSL - new SHA2/AESGCM
> ciphersuites will not be used even when both client and server
> support them.  Also CBC-mode fixes in TLSv1.1 will be missed.
>
> It's a client-side OpenSSL problem and caused indeed
> by following line in fe-secure.c:
>
>         SSL_context = SSL_CTX_new(TLSv1_method());
>
> It's an ugly problem, because TLSv1_method() actually *should* mean
> "TLSv1.0 and higher", but due to problems with various broken
> SSL implementations, it's disabled.
>
> I see various ways it can improve:
>
> 1) OpenSSL guys change default back to TLSv1.0+.
> 2) OpenSSL guys give switch to make TLSv1_method() mean TLSv1.0+.
> 3) libpq starts using TLSv1_2_method() by default.
> 4) libpq will give switch to users to request TLSv1.2.
>
> I see 1) and 3) as unlikely to happen.  As it's not an urgent problem,
> we could watch if 2) happens and go with 4) otherwise.
>
>
> I tried your suggested OP_ALL patch and it does not work.  And it's
> even harmful to use as it disables few security workarounds.
> It's mainly for servers for compat with legacy browsers.
>
> I also tried to clear OP_NO_TLSv1_x to see if there is some default
> OPs in TLSv1_method() that we could change, but that also did not work.
> My conclusion is that currently there is no switch to make TLSv1.0+
> work.  (OpenSSL v1.0.1 / 1.1.0-git).
>