python-peps/peps/pep-0644.rst

PEP: 644
Title: Require OpenSSL 1.1.1 or newer
Author: Christian Heimes <christian@python.org>
Discussions-To: https://discuss.python.org/t/pep-644-require-openssl-1-1-or-newer/5584
Status: Final
Type: Standards Track
Content-Type: text/x-rst
Created: 27-Oct-2020
Python-Version: 3.10
Post-History: 27-Oct-2020, 03-Mar-2021, 17-Mar-2021, 17-Apr-2021
Resolution: https://mail.python.org/archives/list/python-dev@python.org/message/INLCO2EZVQW7R7J2OL6HWVLVU3TQRAZV/


Abstract
========

This PEP proposes for CPython’s standard library to support only OpenSSL
1.1.1 LTS or newer. Support for OpenSSL versions past end-of-lifetime,
incompatible forks, and other TLS libraries are dropped.


Motivation
==========

Python makes use of OpenSSL in ``hashlib``, ``hmac``, and ``ssl`` modules. OpenSSL
provides fast implementations of cryptographic primitives and a full TLS
stack including handling of X.509 certificates. The ``ssl`` module is used by
standard library modules like ``urllib`` and 3rd party modules like ``urllib3``
to implement secure variants of internet protocols. ``pip`` uses the ``ssl``
module to securely download packages from PyPI. Any bug in the ``ssl`` module's
bindings to OpenSSL can lead to a severe security issue.

Over time OpenSSL's public API has evolved and changed. Version 1.0.2
introduced new APIs to verify and match hostnames. OpenSSL 1.1.0 made
internal structs opaque and introduced new APIs that replace direct access of
struct members. Version 3.0.0 will deprecate more APIs due to internal
reorganization that moves cryptographic algorithms out of the core and into
providers. Forks like LibreSSL and BoringSSL have diverged in different
directions.

Currently Python versions 3.6 to 3.9 are compatible with OpenSSL 1.0.2,
1.1.0, and 1.1.1. For the most part Python also works with LibreSSL >= 2.7.1
with some missing features and broken tests.

Due to limited resources and time it becomes increasingly hard to support
multiple versions and forks as well as test and verify correctness. Besides
multiple incompatible APIs there are build time flags,
distribution-specific patches, and local crypto-policy settings that add to
plethora of combinations. On the other hand, the Python core team has only
a couple of domain experts who are familiar with TLS and OpenSSL internals
and even fewer who are active maintainers.

Requiring OpenSSL 1.1.1 would allow us to give the vast majority of users a
better experience, reduce our maintenance overhead and thus free resources
to implement new features. Users would be able to rely on the presence of
new features and consistent behavior, ultimately resulting in a more robust
experience.


Impact
======

OpenSSL 1.1.1 is the default variant and version of OpenSSL on almost all
supported platforms and distributions. It’s also the only version that still
receives security support from upstream [9]_.

No macOS and Windows user will be affected by the deprecation. The python.org
installer and alternative distributions like Conda ship with most recent
OpenSSL version.

As of October 2020 and according to DistroWatch [1]_ most current BSD and
Linux distributions ship with OpenSSL 1.1.1 as well. Some older releases of
long-term support (LTS) and enterprise distributions have older versions of
OpenSSL or LibreSSL. By the time Python 3.10 will be generally available,
several of these distributions will have reached end of lifetime, end of
general support, or moved from LibreSSL to OpenSSL.

Other software has dropped support for OpenSSL 1.0.2 as well. For example,
PyCA cryptography 3.2 (2020-10-25) removed compatibility with OpenSSL 1.0.2.


OpenSSL 1.0.2 LTS
-----------------

released: 2015-02
end of lifetime: 2019-12

OpenSSL 1.0.2 added hostname verification, ALPN support, and elliptic curves.

- CentOS 7 (EOL 2024-06)
- Debian 8 Jessie (EOL 2020-07)
- Linux Mint 18.3 (EOL 2021-04)
- RHEL 7 (full support ends 2019-08, maintenance 2 support ends 2024-06)
- SUSE Enterprise Linux 12-SP5 (general supports ends 2024-10)
- Ubuntu 16.04 LTS / Xenial (general support ends 2021-04)


OpenSSL 1.1.0
-------------

released: 2016-08
end of lifetime: 2019-09

OpenSSL 1.1.0 removed or disabled insecure ciphers by default and added
support for ChaCha20-Poly1305, BLAKE2 (basic features), X25519 and CT. The
majority of structs were made opaque and new APIs were introduced. OpenSSL
1.1.0 is not API compatible with 1.0.2.

- Debian 9 Stretch (security support ended 2020-07, LTS until 2022-06)
- Ubuntu 18.04 LTS / Bionic (general support ends 2023-04)


OpenSSL 1.1.1 LTS
-----------------

released: 2018-08
end of lifetime: 2023-09 (planned)

OpenSSL 1.1.1 added TLS 1.3, SHA-3, X448 and Ed448.

- Alpine (switched back to OpenSSL in 2018 [4]_)
- Arch Linux current
- CentOS 8.0+
- Debian 10 Buster
- Debian 11 Bullseye (ETA 2021-06)
- Fedora 29+
- FreeBSD 11.3+
- Gentoo Linux stable (dropped LibreSSL as alternative in January 2021 [10]_)
- HardenedBSD (switched back to OpenSSL in 2018 [3]_)
- Linux Mint 19.3+
- macOS (python.org installer)
- NetBSD 8.2+
- openSUSE 15.2+
- RHEL 8.0+
- Slackware current
- SUSE Enterprise Linux 15-SP2
- Ubuntu 18.10+
- Ubuntu 20.04 LTS / Focal
- VoidLinux (switched back to OpenSSL in March 2021 [5]_)
- Windows (python.org installer, Conda)

Major CI providers provide images with OpenSSL 1.1.1.

- AppVeyor (with image ``Ubuntu2004``)
- CircleCI (with recent ``cimg/base:stable`` or ``cimg/base:stable-20.04``)
- GitHub Actions (with ``runs-on: ubuntu-20.04``)
- Giblab CI (with Debian Stretch, Ubuntu Focal, CentOS 8, RHEL 8, or Fedora
  runner)
- Packit
- TravisCI (with ``dist: focal``)
- Zuul


OpenSSL 3.0.0
-------------

released: n/a (planned for mid/late 2021)

OpenSSL 3.0.0 is currently under development. Major changes include
relicensing to Apache License 2.0 and a new API for cryptographic algorithms
providers. Most changes are internal refactorings and don’t affect public
APIs. [8]_


LibreSSL
--------

created: 2014-04 (forked from OpenSSL 1.0.1g)

- DragonFly BSD
- Hyperbola GNU/Linux-libre
- OpenBSD
- OpenELEC (discontinued)
- TrueOS (discontinued)

Some distributions like FreeBSD and OPNsense also feature LibreSSL
instead of OpenSSL as non-standard TLS libraries. Gentoo discontinued
LibreSSL as an alternative to OpenSSL in January 2021 [10]_ due to
compatibility issues and little testing.

OpenBSD ports has a port ``security/openssl/1.1`` which is documented as
"[...] is present to provide support for applications which cannot be made
compatible with LibReSSL" [7]_. The package could be used by OpenBSD to
provide a working ssl module.


BoringSSL
---------

created: 2014-06

BoringSSL is Google’s fork of OpenSSL. It’s not intended for general use and
therefore not supported by Python. There are no guarantees of API or ABI
stability.  Vendored copies of BoringSSL are used in Chrome/Chromium browser,
Android, and on Apple platforms [6]_.


Benefits
========

TLS 1.3
-------

OpenSSL 1.1.1 introduced support for the new TLS 1.3 version. The latest
version of the TLS protocol has a faster handshake and is more secure than
the previous versions.

Thread and fork safety
----------------------

Starting with release 1.1.0c, OpenSSL is fully fork and thread safe.
Bindings no longer need any workarounds or additional callbacks to support
multithreading.

SHA-3
-----

Since 1.1.0, OpenSSL ships with SHA-3 and SHAKE implementations.
Python's builtin SHA-3 support is based on the reference implementation. The
internal ``_sha3`` code is fairly large and the resulting shared library close
to 0.5 MB. Python could drop the builtin implementation and rely on OpenSSL's
``libcrypto`` instead.

So far LibreSSL upstream development has refused to add SHA-3 support. [2]_


Compatibility
=============

OpenSSL downstream patches and options
--------------------------------------

OpenSSL features more than 70 configure and build time options in the form
of  ``OPENSSL_NO_*`` macros. Around 60 options affect the presence of features
like cryptographic algorithms and TLS versions. Some distributions apply
patches to alter settings. Furthermore, default values for settings like
security level, ciphers, TLS version range, and signature algorithms can
be set in OpenSSL config file.

The Python core team lacks resources to test all possible combinations.
This PEP proposes that Python only supports OpenSSL builds that have
standard features enabled. Vendors shall disable deprecated or insecure
algorithms and TLS versions with build time options like
``OPENSSL_NO_TLS1_1_METHOD`` or OpenSSL config options like
``MinProtocol = TLSv1.2``.

Python assumes that OpenSSL is built with

- hashlib’s default algorithms such as MD5, SHA-1, SHA-2 family,
  SHA-3/SHAKE family, BLAKE2
- TLS 1.2 and TLS 1.3 protocols
- current key agreement, signature, and encryption algorithms for TLS 1.2
  and 1.3 (ECDH, RSA, ECDSA, Curve25519, AES, Poly1309-ChaCha20, ...)
- threading, file I/O, socket I/O, and error messages

Weak algorithms (MD5, SHA-1 signatures) and short keys (RSA < 2024 bits) may
be disabled at runtime. Algorithms may also be blocked when they are
disabled by a crypto policy such as FIPS. The PEP is not more specific on
purpose to give room for new features as well as countermeasures against
vulnerabilities. As a rule of thumb, Python should be able to connect to
PyPI and the test suite should pass.

LibreSSL support
----------------

LibreSSL is a fork of OpenSSL. The fork was created off OpenSSL 1.0.1g by
members of the OpenBSD team in 2014 in light of the heartbleed vulnerability.
Since its inception several features deemed problematic or insecure were
removed or replaced (SSL 2.0, SSL 3.0, improved CPRNG) or backported
from OpenSSL and BoringSSL.

At the moment LibreSSL is not fully API compatible with OpenSSL 1.1.1. The
latest release LibreSSL 3.3.2 is missing features and behaves differently
in some cases. Mentionable missing or incompatible features include

- SHA-3, SHAKE, BLAKE2
- ``SSL_CERT_*`` environment variables
- security level APIs
- session handling APIs
- key logging API
- verified cert chain APIs
- OPENSSL_VERSION macro

This PEP proposed to remove any and all LibreSSL related workarounds from
Python. In the future Python will not actively prohibit LibreSSL support
with configure and compile time checks. But Python will not accept patches
that add non-trivial workarounds or disable tests either.


BoringSSL
---------

There are currently no plans to support BoringSSL.


Rejected Ideas
==============

Formalize supported OpenSSL versions
------------------------------------

This PEP does not provide a set of formal rules and conditions under which
an OpenSSL version is supported.

In general Python aims to be compatible with commonly used and officially
supported OpenSSL versions. Patch releases of Python may not be compatible
with new major releases of OpenSSL. Users should not expect that a new major
or minor  release of Python works with an OpenSSL version that is past its
end-of-lifetime. Python core development may backport fixes for new releases
or extend compatibility with EOLed releases as we see fit.

The new ABI stability and LTS policies of OpenSSL [9]_ should help, too.

Keep support for OpenSSL 1.1.0
------------------------------

It was suggested to keep support for OpenSSL 1.1.0 for compatibility with
Debian 9 (Stretch). The proposal was rejected since it would complicated code
cleanup and testing. Stretch is already out of regular security support and
close to end of long-term support. By the time of Python 3.10 final release,
Debian Buster and Debian Bullseye will be available.

Instead Python 3.10 will gain additional documentation and a new
``configure`` option ``--with-openssl-rpath=auto`` to simplify use of custom
OpenSSL builds [11]_.


Backwards Compatibility
=======================

Python 3.10 will no longer support TLS/SSL and fast hashing on platforms
with OpenSSL 1.0.2 or LibreSSL. The first draft of this PEP was published at
the beginning of the 3.10 release cycles to give vendors like Linux
distributors or CI providers sufficient time to plan.

Python's internal copy of the *Keccak Code Package* and the internal
``_sha3`` module will be removed. This will reduce source code size by
about 280kB and code size by roughly 0.5MB. The ``hashlib`` will solely rely
on OpenSSL's SHA-3 implementation. SHA-3 and SHAKE will no longer be available
without OpenSSL.


Disclaimer and special thanks
=============================

The author of this PEP is a contributor to OpenSSL project and employed by
a major Linux distributor that uses OpenSSL.

Thanks to Alex Gaynor, Gregory P. Smith, Nathaniel J. Smith, Paul Kehrer,
and Seth Larson for their review and feedback on the initial draft.


References
==========

.. [1] https://distrowatch.com/
.. [2] https://github.com/libressl-portable/portable/issues/455
.. [3] https://hardenedbsd.org/article/shawn-webb/2018-04-30/hardenedbsd-switching-back-openssl
.. [4] https://lists.alpinelinux.org/~alpine/devel/%3CCA%2BT2pCGFeh30aEi43hAvJ3yoHBijABy_U62wfjhVmf3FmbNUUg%40mail.gmail.com%3E
.. [5] https://voidlinux.org/news/2021/02/OpenSSL.html
.. [6] https://forums.swift.org/t/rfc-moving-swiftnio-ssl-to-boringssl/18280
.. [7] https://openports.se/security/openssl/1.1
.. [8] https://www.openssl.org/docs/OpenSSL300Design.html
.. [9] https://www.openssl.org/policies/releasestrat.html
.. [10] https://www.gentoo.org/support/news-items/2021-01-05-libressl-support-discontinued.html
.. [11] https://bugs.python.org/issue43466


Copyright
=========

This document is placed in the public domain or under the
CC0-1.0-Universal license, whichever is more permissive.