ssl.create_default_context()

[tiran]

BPO	19689
Nosy	@gvanrossum, @pitrou, @tiran
Files	ssl_create_default_context.patch ssl_create_default_context2.patch ssl_create_default_context3.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/tiran'
closed_at = <Date 2014-10-04.21:18:44.487>
created_at = <Date 2013-11-22.03:22:24.405>
labels = ['type-feature', 'library']
title = 'ssl.create_default_context()'
updated_at = <Date 2014-10-04.21:18:44.486>
user = 'https://github.com/tiran'

bugs.python.org fields:

activity = <Date 2014-10-04.21:18:44.486>
actor = 'pitrou'
assignee = 'christian.heimes'
closed = True
closed_date = <Date 2014-10-04.21:18:44.487>
closer = 'pitrou'
components = ['Library (Lib)']
creation = <Date 2013-11-22.03:22:24.405>
creator = 'christian.heimes'
dependencies = []
files = ['32770', '32778', '32799']
hgrepos = []
issue_num = 19689
keywords = ['patch']
message_count = 16.0
messages = ['203718', '203757', '203758', '203761', '203765', '203768', '203770', '203774', '203812', '203813', '203816', '203842', '203844', '204031', '204039', '213010']
nosy_count = 5.0
nosy_names = ['gvanrossum', 'janssen', 'pitrou', 'christian.heimes', 'python-dev']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue19689'
versions = ['Python 3.4']

[tiran]

A few weeks ago I suggested the addition of ssl.create_default_context() to the stdlib. The patch implements my proposal. It replaces code in several of modules with one central function. The patch also removes ssl.wrap_socket() in favor for a SSLContext object.

As soon as bpo-19292 gets accepted I'll add the additional keyword argument "purpose=None" to the arguments of ssl.create_default_context() in order to load the default system certs::

if purpose is not None and verify_mode != CERT_NONE:
    context.load_default_certs(purpose)

[pitrou]

A few weeks ago I suggested the addition of
ssl.create_default_context() to the stdlib. The patch implements my
proposal. It replaces code in several of modules with one central
function. The patch also removes ssl.wrap_socket() in favor for a
SSLContext object.

Can you call it create_default_client_context() (a bit long) and/or
stress that it's for client use?

Or will it be ok for server purposes too, i.e. do you promise that it'll
never get CERT_REQUIRED by default?

[tiran]

Good point!

We need a purpose flag anyway in order to load the appropriate root CA certs. The purpose flag can be used for purpose-specific verify mode:

SERVER_AUTH = _ASN1Object('1.3.6.1.5.5.7.3.1')
CLIENT_AUTH = _ASN1Object('1.3.6.1.5.5.7.3.2')

    if isinstance(purpose, str):
        purpose = _ASN1Object.fromname(purpose)
    if verify_mode is None:
        if purpose == SERVER_AUTH:
            # authenticate a TLS web server (for client sockets). The default 
            # setting may change in the future.
            verify_mode = CERT_NONE
        elif purpose == CLIENT_AUTH:
            # authenticate a TLS web client (for server sockets). The default
            # setting is guaranteed to be stable and will never change.
            verify_mode = CERT_NONE
        else:
            # other (code signing, S/MIME, IPSEC, ...), default may change.
            verify_mode = CERT_NONE
    context.verify_mode = verify_mode

[pitrou]

SERVER_AUTH = _ASN1Object('1.3.6.1.5.5.7.3.1')
CLIENT_AUTH = _ASN1Object('1.3.6.1.5.5.7.3.2')

That's a bit ugly. How about an enum?

[tiran]

In my opinion enums are for a closed batch of known entities. There are at least 20-30 purpose flags, maybe more. Everybody is allowed to define their own OIDs, too.

[pitrou]

In my opinion enums are for a closed batch of known entities. There
are at least 20-30 purpose flags, maybe more. Everybody is allowed to
define their own OIDs, too.

Well, how many purposes are we going to expose? I don't think users
should know what ASN1 objects are, and a nice repr() is really useful.

(but there's no reason an enum cannot have 20 or 30 members)

[tiran]

The objects already have a (more or less) nice representation:

>>> ssl._ASN1Object.fromname("1.3.6.1.5.5.7.3.1")
_ASN1Object(nid=129, shortname='serverAuth', longname='TLS Web Server Authentication', oid='1.3.6.1.5.5.7.3.1')
>>> ssl._ASN1Object.fromname("1.3.6.1.5.5.7.3.2")
_ASN1Object(nid=130, shortname='clientAuth', longname='TLS Web Client Authentication', oid='1.3.6.1.5.5.7.3.2')
>>> ssl._ASN1Object.fromname("1.3.6.1.5.5.7.3.8")
_ASN1Object(nid=133, shortname='timeStamping', longname='Time Stamping', oid='1.3.6.1.5.5.7.3.8')

[pitrou]

Ok. Note that as long as they aren't actually passed to OpenSSL, they don't need to be ASN1 objects at all, i.e. if it's only a parameter to create_default_context(), it can perfectly well be a str or enum.

[tiran]

New patch with enum and more cleanups.

I'd like to explain the rationals for the purpose argument in create_default_context and the ASN1Object thing. There are multiple things involved here. First of all a certificate may have key usage and extended key usage OIDs in its X509v3 extensions. OpenSSL already checks them according to its mode.

The purpose is also required to load the correct set of certs from a certificate provider (e.g. Windows cert store, Mozilla NSS certdata, Apple's keystore). The system or user can impose additional restrictions for certificates, e.g. disable a cert for TLS web server auth although the X.509 struct specifies 1.3.6.1.5.5.7.3.1 in its X509v3 extensions. NSS certdata also contains invalid certificates or certificates that are not suitable for server auth although the cert claims it.

In order to load only trusted certs for a purpose the API needs a purpose flag (usually an OID or a NID). Most Linux users have never seen this differentiation because /etc/ssl/certs/ either contains only server auth certs or their distributions screw up, See https://bugs.launchpad.net/ubuntu/+source/ca-certificates/+bug/1207004 or http://www.egenix.com/company/news/eGenix-pyOpenSSL-Distribution-0.13.2.1.0.1.5.html

[tiran]

More links:

https://www.imperialviolet.org/2012/01/30/mozillaroots.html
curl/curl@51f0b798fa
https://github.com/kennethreitz/requests/issues/1659

[pitrou]

Ok, so I still have a couple of issues with the proposed API:

• if its purpose is to create a *default* context, create_default_context() shouldn't have that many arguments. The nice thing with contexts is that you can change their parameters later... So basically the function signature should be:

  create_default_context(purpose, *, cafile, cadata, capath)

Also, the default for "purpose" should probably be serverAuth.

• "PurposeEKU" is cryptic, please simply "Purpose" or "CertPurpose".

[tiran]

Antoine and I have agreed upon a slightly different API. I'm going to split it up into one public API that creates a best practice context and one internal stdlib API to unify all places that deals with SSL sockets.

AP:
how about we use more strict and modern settings for the public API? TLSv1, no insecure stuff like RC4, MD5, DSS etc. https://hynek.me/articles/hardening-your-web-servers-ssl-ciphers/

[pitrou]

how about we use more strict and modern settings for the public API?
TLSv1, no insecure stuff like RC4, MD5, DSS etc.
https://hynek.me/articles/hardening-your-web-servers-ssl-ciphers/

Fine, but I'd like to see something more open-ended for the ciphers
string. e.g.
'HIGH:!ADH:!AECDH:!MD5:!DSS:!aNULL:!eNULL:!LOW:!EXPORT:!SSLv2' ?

[tiran]

The patch implements HIGH:!aNULL:!RC4:!DSS

HIGH already covers !MD5:!EXPORT:!NULL:!SSLv2 and more

[python-dev]

New changeset 63df21e74c65 by Christian Heimes in branch 'default':
Issue bpo-19689: Add ssl.create_default_context() factory function. It creates
http://hg.python.org/cpython/rev/63df21e74c65

[python-dev]

New changeset 8b4b6609cd31 by R David Murray in branch 'default':
whatsnew: ssl.create_default_context (bpo-19689).
http://hg.python.org/cpython/rev/8b4b6609cd31