On Sat, 16 Nov 2013 14:11:33 +0100 Julien Danjou <julien@danjou.info> wrote: 

JD> On Sat, Nov 16 2013, Vincent Bernat wrote:
>> In the same way as for whitelisting, default verification options should
>> be a variable with possibility to override it by using the appropriate
>> option of `gnutls-negotiate`.

OK, I worked on this.

>> Verification options could be:
>> 
>> - `expired-certificate`
>> - `revoked-certificate`
>> - `untrusted-certificate`
>> - `hostname-mismatch`

I'm not sure this granularity is necessary.  I just have :trustfiles and
:hostname as options right now.  Anyone else with an opinion?  Note that
you can also specify verification flags.  From the `gnutls-boot' docs:

VERIFY-FLAGS is a numeric OR of verification flags only for
`gnutls-x509pki' connections.  See GnuTLS' x509.h for details;
here's a recent version of the list.

    GNUTLS_VERIFY_DISABLE_CA_SIGN = 1,
    GNUTLS_VERIFY_ALLOW_X509_V1_CA_CRT = 2,
    GNUTLS_VERIFY_DO_NOT_ALLOW_SAME = 4,
    GNUTLS_VERIFY_ALLOW_ANY_X509_V1_CA_CRT = 8,
    GNUTLS_VERIFY_ALLOW_SIGN_RSA_MD2 = 16,
    GNUTLS_VERIFY_ALLOW_SIGN_RSA_MD5 = 32,
    GNUTLS_VERIFY_DISABLE_TIME_CHECKS = 64,
    GNUTLS_VERIFY_DISABLE_TRUSTED_TIME_CHECKS = 128,
    GNUTLS_VERIFY_DO_NOT_ALLOW_X509_V1_CA_CRT = 256

It must be omitted, a number, or nil; if omitted or nil it
defaults to GNUTLS_VERIFY_ALLOW_X509_V1_CA_CRT.

JD> I think this is a really good idea and I'm waiting for that for a long
JD> time. Since I don't have time to do this myself currently, consider this
JD> message as a strong support. \o/

JD> I'd be happy to help and test as far as I can.

First patch attached here against Emacs trunk.  It removes
`:verify-hostname-error' everywhere, adds a new `gnutls-verify-error'
defcustom and uses it in `gnutls-boot', and allows the user to match the
hostname to a regexp and provide verification flags for each such
regexp.

To make verification errors abort GnuTLS negotiations, the
`gnutls-verify-error' simply needs to be t (it's nil in this patch).

My concern is that suddenly connections will start failing for our users
and bug reports will flow, and I don't have time to explain to everyone
why their self-signed certificates need exceptions.  This can be really,
really annoying.  But logging in *Messages* is not very useful either,
users don't read it.  So what's the right thing?  How about a default
behavior of flashing a warning, then sit-for 3 seconds?  A hard error
can be optional but not the default.

I'm also not sure I like the look and feel of the `gnutls-verify-error'
defcustom.  It's kind of awkward.

I'd like to get this done before the Emacs code freeze next week or so.
Please give me your opinions and test the code.

Ted