Support NSSDB in Python API
- 1 [Support NSSDB in PKI's Python API]
- 1.1 Overview
- 1.2 Associated Bugs and Tickets
- 1.3 Use Cases
- 1.4 Operating System Platforms and Architectures
- 1.5 Technology Stack
- 1.6 Design
- 1.7 Implementation
- 1.8 Major configuration options and enablement
- 1.9 Cloning
- 1.10 Open Questions
- 1.11 Alternative design
- 1.12 Further things to consider
- 1.13 Updates and Upgrades
- 1.14 Tests
- 1.15 Dependencies
- 1.16 Packages
- 1.17 External Impact
- 1.18 History
[Support NSSDB in PKI's Python API]
Draft, work in progress!
FreeIPA clients use X.509 client certs to authenticate themselves against the server. FreeIPA stores certificates securely in a NSSDB and depends on python-nss for secure connections and client certificates handling. Internally it uses a low level httplib interface with a custom connection factory. [XXX: verify]
Dogtag API uses python-requests for HTTP connections. Python-requests also known as "HTTP for Humans" because it provides a modern, feature-rich yet easy to use API for HTTP. However it depends on Python's SSL stack and OpenSSL, which are both unable to retrieve certificates from a NSSDB directly. Further more python-requests <= 2.7 and the underlying library urllib3 only expose a subset of Python's SSL module, which is already limited all by itself.
As a consequence and in order to make authenticated requests Dogtag's Python API has to keep the client certs private key in an unencrypted PEM file on the file system. Python's SSL module can neither interface NSSDB directly nor load certs from memory. Requests doesn't support a password callback for encrypted certificates although it is supported by Python's SSL module. The need for an unencrypted copy of the client certificate adds an extra synchronisation step and imposes a security risk as well.
This design document explores possible ways to directly and securely interface NSSDB from Dogtag API.
Associated Bugs and Tickets
Walk through one or more full examples of how the feature will be used. These should not all be the simplest cases.
Operating System Platforms and Architectures
The feature will be developed and initially tested on Linux (Fedora 22). It should run on other Linux platforms easily.
Python-requests is build on top of a deep technology stack. This propals plan to make the stack even deeper. This chapter explains the relationship between the involved libraries.
The requests package provides the public interface. Internally it contains 3rd party package such as urllib3 and a copy of Mozilla's trust store.
urllib3 is used as the underlying library for connection handling and HTTP parsing. It's a redesign of Python's urllib and urllib2 modules (hence the name) with a more modern design. It has elaborate features such a connection pool for HTTP/1.1 keep-alive. urllib3 uses Python's socket module for network connections and Python's [https://docs.python.org/3/library/ssl.html ssl] module for TLS/SSL by default. It can use PyOpenSSL as an alternative SSL library, too. As of now urllib3 1.10 does not support advanced options of both ssl and PyOpenSSL, for instance password callbacks for password protected private keys.
PyOpenSSL is an alternative TLS/SSL package with a different API. The basic concepts are similar to Python's standard library but method names and arguments are different. It's possible to adapt PyOpenSSL to a mostly compatible implementation with e.g. hyper's ssl_compat
PyOpenSSL used to be written in plain C, but it was rewritten a while ago. Nowadays it relies on PyCA's cryptography and CFFI (C Foreign Function Interface for Python). CFFI simplifies development and ensure compatibility with CPython and PyPy without extra work. That means improvements for PyOpenSSL are available in PyPy!
Both the ssl module in Python's standard library and PyOpenSSL rely on OpenSSL for TLS/SSL protocol, X.509 handling and cryptographic primitives. OpenSSL is able to load client certs in DER and PEM format from files or in-memory buffers. It can't interface with NSSDB directly.
NSS (Mozilla's Network Security Services) is an alternative TLS/SSL library that is used in amongst others Firefox and Red Hat Directory Server. NSSDB (aka CertDB) provides the certificate store for trust anchors, client certs and other X.509 certificates.
python-nss is a Python wrapper for NSS.
In order to provide NSSDB support in python-requests OpenSSL must be able look up client certificates in NSSDB. Several upstream packages must be altered and improved to achieve that goal. Since Python's ssl module is feature-locked to new minor releases of Python, this proposal only focuses on PyOpenSSL as TLS library. Even if the necessary improvements are implemented and later accepted by upstream, it would take more than 18 months until the release of Python 3.6.0 before the mandatory callback is available.
The idea here boils down to a SSL context with a client cert request callback, that fetches the client certificate from NSSDB during the SSL handshake and returns it to the server. In OpenSSL a SSL context acts as a central configuration and session instance for one or more SSL connection. Beside options like TLS protocol, ciphers, verification and validation flags it holds trusted root CAs, certificates, callback functions and many more. A SSL context contains several hooks for client certificate requests.
In both Python's ssl module and PyOpenSSL SSL context objects play the same role as configuration instance. Contexts also act as factories which wrap BSD socket connections into a SSL-aware connections. Libraries usually create a TCP/IP connection first and then upgrade the connection to a SSL connection.
All network related Python standard library modules like urllib or smtplib take an optional SSLContext object as an argument. As of now neither urllib3 nor python-requests support a context argument. Upstream is going to implement the feature soonish.
The SSLContext class of Python's ssl module and PyOpenSSL's [https://pyopenssl.readthedocs.org/en/latest/api/ssl.html#context-objects Context] type have a different and incompatible API. urllib3 only wants to support a Python stdlib compatible SSLContext object. The PyOpenSSL Context type must be wrapped in a SSLContext compatible interface.
python-nss needs to support a lookup method for client certs that is able to handle information from PyOpenSSL.
Finally Dogtag needs to set up all necessary bits and pieces.
from OpenSSL.crypto import load_certificate, FILETYPE_ASN1 from OpenSSL.SSL import Context from nss import nss import requests certdb = nss.get_default_certdb() def client_cert_cb(sslcon): for cert in in sslcon.get_peer_cert_chain(): issuer = cert.get_issuer().der() client_cert = certdb.find_client_cert_by_issuer(issuer) if client_cert is not None: x509 = load_certificate(FILETYPE_ASN1, client_cert.der()) sslcon.add_client_cert(x509) break pyctx = Context(OpenSSL.SSL.SSLv23_METHOD) pyctx.set_client_cert_cb(client_cert_cb) # more configuration ctx = SSLContextAdapter(pyctx) requests.get('https://dogtag.example.org/api', ssl_context=ctx)
PyCA Cryptography has to wrap and expose additional functions. First of all we need one or both of the callback setters in order to install a callback function in a SSL context:
For the more generic SSL_CTX_set_cert_cb() the client public and private key must be added to the SSL connection instead of the SSL context. Cryptography just exposes the SSL_CTX variants of the functions yet, so these four function must be provided, too. The ASN1 variants are probably easier to use in order to add certs from NSSDB.
- SSL_use_certificate( )
Finally this check function can become in handy:
PyOpenSSL has to include these new function and add them to the appropriate classes.
pull request #570 for ssl_context argument is work in progress.
python-requests needs to support a ssl_context argument, too.
PyOpenSSL to stdlib SSL adapter
Cory Benfield's HTTP/2 library hyper has a ssl_compat module. He has expressed his interest in a standalone compatibility layer.
TBD, python-nss may need to provide additional functions in order to get a client cert by issuer DN for purpose TLS Web Client Authentication.
Dogtag API is where it all comes together.
1) Dogtag has to create and carefully configure a PyOpenSSL Context object and wrap it into a SSLContext adapter. Options (TLS protocol version, ciphers, trust anchors) should probably be taken from a config file with sensible secure settings applied. For performance reasons the SSL context should be cached. A single instance make use of SSL session resumption, too.
2) Dogtag API needs to open the correct NSS CertDB and install a callback like shown in the pseudo code example in the design section.
3) All HTTP API calls with requests must apply the SSL context object as ssl_context argument.
Major configuration options and enablement
Any configuration options? Any commands to enable/disable the feature or turn on/off its parts?
Any impact on cloning?
- Should it be possible to specify a different NSS' CertDB than the default
- How should the API deal with a password callback for encrypted client
certs in NSSDB?
- What is the correct way to lookup client certs in NSSDB? How to deal with
multiple matching client certs?
OpenSSL has a little known feature called ENGINE cryptographic modules. An engine can replace or offload certain functions of OpenSSL, e.g. run RSA on a hardware accelerator, get strong random data from a hardware CPRNG or load a private key from a smart card. Engines can be compiled into OpenSSL as well as loaded dynamically at runtime.
Further more there exists an undocumented feature called ENGINE_set_load_ssl_client_cert_function() in the public engine interface. This engine function hooks into ssl/s3_clnt.c:ssl_do_client_cert_cb() right before the SSL_CTX_set_client_cert_cb() callback. OpenSSL's CAPI engine for Windows uses the feature in capi_load_ssl_client_cert() to provide an engine that fetches client certificates from Windows' cert store. By default a SSL_CTX doesn't use an engine for client cert lookups. The engine must be installed with SSL_CTX_set_client_cert_engine().
We might want to consider this approach because it has some benefits. An engine hook would work with Python's ssl module, too. A pure C implementation would even work for every program that uses OpenSSL.
Further things to consider
Implement X509_LOOKUP to retrieve trust anchors from NSSDB? A NSS-based X509_LOOKUP could eliminate the need for /etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem. It's out of scope for this proposal but it might be worth investigation in the future.
Updates and Upgrades
Any impact on updates and upgrades?
- A new version of python-requests with support for a ssl_context argument.
- A new version of urllib3 with support for a ssl_context argument.
- An adapter that is able to adapt PyOpenSSL's Context and Connection instances into Python stdlib compatible APIs.
- A new version of PyOpenSSL with support for a client cert callback and the injection of client certs into a connection.
- A new version of Cryptography with the necessary infrastructure for PyOpenSSL.
Impact on other development teams and components?
ORIGINAL DESIGN DATE: [Provide the original design date in 'Month DD, YYYY' format (e. g. - September 5, 2013).]
Don't use python-requests
Ticket #1360 listed an alternative approach. Instead of interfacing NSSDB from python-requests Dogtag could also be refactored to not use python-requests but rather a custom library. The alternative is not desired. Requests is a convenient solution and speeds up development considerable. Dogtag can also benefit from more advanced features of python-requests like connection pooling or future features like HTTP/2.
Besides the accompanying improvements in python-requests and other 3rd party components are mutually benefitial for Dogtag, FreeIPA and other Open Source projects throughout and outside Red Hat.
Replace OpenSSL with NSS as TLS library
PyOpenSSL has several vantages compaerd to python-nss. In general we like to benefits from the improvements and features of PyCA Cryptography like PyPy support. NSPR's abstraction layer are another issue. It is incompatible with Python's sockets and asyncio because NSPR uses its own I/O layer. In theory the underlying BSD socket could be imported into a NSPR with PR ImportTCPSocket but even the documentation warns about compatibility issues. Finally OpenSSL and PyOpenSSL are better understood and maintained by more people than NSS.