Merge lp:~jelmer/bzr/urllib-verifies-ssl-certs into lp:bzr

[Jelmer Vernooij]

Add support for verifying SSL certificates when using urllib.

This branch does actually work, but it's got some open ends:

* The location with the CAs to trust differs per system. Rather than hardcoding
  a list, I think we should have a configuration option that allows the user
  to specify what CA list they want to verify against.

  This option could then also be pointed at a user-specific list
  (rather than a system wide list). And packagers can put in the default
  for their distribution.

  Where do I obtain this option from, which config stack ?

* We probably want some way to specify that self-signed certificates are ok.
  Probably also a new option? We should probably still warn the user
  when a self-signed certificate is being trusted.

* There are no additional tests. There probably should be some, at least for
  e.g. the match_hostname helper.

* I have the feeling that I'm missing something. This was way too easy.
  What is it?

[Martin Pool (mbp)]

On 4 November 2011 14:03, Jelmer Vernooij <email address hidden> wrote:
> Jelmer Vernooij has proposed merging lp:~jelmer/bzr/urllib-verifies-ssl-certs into lp:bzr.
>
> Requested reviews:
>  bzr-core (bzr-core)
> Related bugs:
>  Bug #125055 in Bazaar: "defaulting to pycurl doesn't make sense anymore"
>  https://bugs.launchpad.net/bzr/+bug/125055
>  Bug #651161 in Bazaar: "bzr fails to verify ssl validity in https connections - by default --> as pycurl isn't a dep only a suggestion "
>  https://bugs.launchpad.net/bzr/+bug/651161
>
> For more details, see:
> https://code.launchpad.net/~jelmer/bzr/urllib-verifies-ssl-certs/+merge/81227
>
> Add support for verifying SSL certificates when using urllib.

That's great, thanks.

> This branch does actually work, but it's got some open ends:
>
> * The location with the CAs to trust differs per system. Rather than hardcoding
>  a list, I think we should have a configuration option that allows the user
>  to specify what CA list they want to verify against.
>
>  This option could then also be pointed at a user-specific list
>  (rather than a system wide list). And packagers can put in the default
>  for their distribution.
>
>  Where do I obtain this option from, which config stack ?

[needsfixing] I think the global stack would be a good place to start
- it would seem like a kind of contrived case to want different values
on a smaller scope. We can always change it later.

> * We probably want some way to specify that self-signed certificates are ok.
>  Probably also a new option? We should probably still warn the user
>  when a self-signed certificate is being trusted.

+1 to a new option and to always warning.

Though perhaps it would be equivalent and easier to just have an
option saying "don't do validation."

> * There are no additional tests. There probably should be some, at least for
>  e.g. the match_hostname helper.

It would be nice, if it's not too hard, to actually test against an
untrusted cert, or at least to do monkeypatching to make it look like
we got an invalid cert.

[needsfixing] I'd like a unit test for match_hostname covering the
various cases.

> * I have the feeling that I'm missing something. This was way too easy.
>  What is it?

I haven't read the rfc but it looks plausible. You could wait for a
review from vila, who has probably touched this most, or robert who is
an http nerd.

I wonder if this needs a change to Windows or Mac packaging to make
sure the certs are there. Or, possibly we can get them from the
system's keyring in some way.

[Vincent Ladeuil (vila)]

> Add support for verifying SSL certificates when using urllib.

\o/

>
> This branch does actually work, but it's got some open ends:
>
> * The location with the CAs to trust differs per system. Rather than hardcoding
> a list, I think we should have a configuration option that allows the user
> to specify what CA list they want to verify against.

pycurl already handles that for windows but getting it right in the general
case... sounds like a packaging issue :-/

>
> This option could then also be pointed at a user-specific list
> (rather than a system wide list). And packagers can put in the default
> for their distribution.
>
> Where do I obtain this option from, which config stack ?

I'd go with global to start with specifying the default value at
registration time if possible.

Arguably it's a system-wide config file that should be used, making packager
life easier.

>
> * We probably want some way to specify that self-signed certificates are ok.
> Probably also a new option? We should probably still warn the user
> when a self-signed certificate is being trusted.

authentication.conf is the right place for that and the option is even
documented but not implemented ;)

>
> * There are no additional tests. There probably should be some, at least for
> e.g. the match_hostname helper.

Yup. There are already https tests for pycurl that should be activated now
(<broken record> hey, did I mention the 'scenarios' attribute is cool for
that ? </broken record> ;)

>
> * I have the feeling that I'm missing something. This was way too easy.
> What is it?

Hehe.

The last time I worked on this area we were still supporting python-2.4/2.5
where https wasn't supported. Now, you also found the relevant code in
python-3.2 so... that certainly simplify things *a lot* :)

This wasn't a full review but as I understand it you weren't requesting one
either ;)

As for tests, if you look at the pycurl ones you'll certainly find that most
of the needed infrastructure is there to build certificates for special
purposes (self-signed, obsolete, etc) and easily setup an https server.

99 + SSLSocket.getpeercert()) matches the *hostname*. RFC 2818 rules
100 + are mostly followed, but IP addresses are not accepted for *hostname*.

Hmm, IP addresses not accepted ? I bet our tests won't like that much.

pycurl itself was not open enough to test for these weird certificates AFAIR
but it's also a robust implementation so I wasn't concerned, but if we make
urllib the default one, we have to make sure we're making the right checks.

I'll need to dig a bit more to find back the relevant discussions about this
topic.

Also, defaulting to urllib is fine but the documentation probably needs to
be updated (created even ?) to make it clear that we're not defaulting to
pycurl. We had issues in the past (including failures to check lp
certificate) which had led to the actual defaults (urllib for http, pycurl
for https), let's make the switch crystal clear.

141 def connect_to_origin(self):
142 - ssl_sock = _ssl_wrap_socket(self.sock, self.key_file, self.cert_file)
143 + ssl_sock = ssl.wrap_socket(self.sock, self.key_file, self.cert_file,
144 + ...

[Martin Packman (gz)]

I think there's general agreement this is great, and with the reqirement of Python 2.6 something that's now acheivable and should probably get into the next bzr release.

It absolutely needs some kind of automated testing though, and likely an option for disabling checks, I've used `curl -k` and `wget --no-check-certificate` enough to know I'd be annoyed if that option wasn't there.

review: Needs Fixing

=== modified file 'bzrlib/config.py'
--- bzrlib/config.py	2011-12-19 16:41:49 +0000
+++ bzrlib/config.py	2011-12-20 09:40:32 +0000
@@ -1662,7 +1662,6 @@
         return value.decode(osutils.get_user_encoding())
     return unicode_str
 
-
 def _auto_user_id():
     """Calculate automatic user identification.
 
@@ -2761,6 +2760,13 @@
            help="If we wait for a new request from a client for more than"
                 " X seconds, consider the client idle, and hangup."))
 
+option_registry.register_lazy('ssl.ca_certs',
+    'bzrlib.transport.http._urllib2_wrappers', 'opt_ssl_ca_certs')
+
+option_registry.register_lazy('ssl.cert_reqs',
+    'bzrlib.transport.http._urllib2_wrappers', 'opt_ssl_cert_reqs')
+
+
 
 class Section(object):
     """A section defines a dict of option name => value.
 
=== modified file 'bzrlib/errors.py'
--- bzrlib/errors.py	2011-12-19 13:23:58 +0000
+++ bzrlib/errors.py	2011-12-20 09:40:32 +0000
@@ -1655,6 +1655,14 @@
         TransportError.__init__(self, msg, orig_error=orig_error)
 
 
+class CertificateError(TransportError):
+
+    _fmt = "Certificate error: %(error)s"
+
+    def __init__(self, error):
+        self.error = error
+
+
 class InvalidHttpRange(InvalidHttpResponse):
 
     _fmt = "Invalid http range %(range)r for %(path)s: %(msg)s"
 
=== modified file 'bzrlib/tests/test_http.py'
--- bzrlib/tests/test_http.py	2011-10-12 16:00:13 +0000
+++ bzrlib/tests/test_http.py	2011-12-20 09:40:32 +0000
@@ -32,7 +32,6 @@
 import bzrlib
 from bzrlib import (
     bzrdir,
-    cethread,
     config,
     debug,
     errors,
 
=== modified file 'bzrlib/transport/__init__.py'
--- bzrlib/transport/__init__.py	2011-12-19 10:58:39 +0000
+++ bzrlib/transport/__init__.py	2011-12-20 09:40:32 +0000
@@ -1765,15 +1765,15 @@
                  help="Read-only access of branches exported on the web.")
 register_transport_proto('https://',
             help="Read-only access of branches exported on the web using SSL.")
-# The default http implementation is urllib, but https is pycurl if available
+# The default http implementation is urllib
 register_lazy_transport('http://', 'bzrlib.transport.http._pycurl',
                         'PyCurlTransport')
+register_lazy_transport('https://', 'bzrlib.transport.http._pycurl',
+                        'PyCurlTransport')
 register_lazy_transport('http://', 'bzrlib.transport.http._urllib',
                         'HttpTransport_urllib')
 register_lazy_transport('https://', 'bzrlib.transport.http._urllib',
                         'HttpTransport_urllib')
-register_lazy_transport('https://', 'bzrlib.transport.http._pycurl',
-                        'PyCurlTransport')
 
 register_transport_proto('ftp://', help="Access using passive FTP.")
 register_lazy_transport('ftp://', 'bzrlib.transport.ftp', 'FtpTransport')
 
=== modified file 'bzrlib/transport/http/_urllib2_wrappers.py'
--- bzrlib/transport/http/_urllib2_wrappers.py	2011-12-19 13:23:58 +0000
+++ bzrlib/transport/http/_urllib2_wrappers.py	2011-12-20 09:40:32 +0000
@@ -63,12 +63,53 @@
     config,
     debug,
     errors,
+    lazy_import,
     osutils,
     trace,
     transport,
     urlutils,
     )
-
+lazy_import.lazy_import(globals(), """
+import ssl
+""")
+
+
+def default_ca_certs():
+    """Find the default file with ca certificates."""
+    return u"/etc/ssl/certs/ca-certificates.crt"
+
+def default_cert_reqs():
+    return u"required"
+
+def cert_reqs_from_store(unicode_str):
+    import ssl
+    try:
+        return {
+            "required": ssl.CERT_REQUIRED,
+            "optional": ssl.CERT_OPTIONAL,
+            "none": ssl.CERT_NONE
+            }[unicode_str]
+    except KeyError:
+        raise ValueError("invalid value %s" % unicode_str)
+
+
+opt_ssl_ca_certs = config.Option('ssl.ca_certs',
+        default=default_ca_certs,
+        help="""\
+Path to certification authority certificates to trust.
+""")
+
+opt_ssl_cert_reqs = config.Option('ssl.cert_reqs',
+        default=default_cert_reqs,
+        from_unicode=cert_reqs_from_store,
+        help="""\
+Whether to require a certificate from the remote side. (default:required)
+
+Possible values:
+ * none: certificates ignored
+ * optional: Certificates not required, but validated if provided
+ * required: Certificates required, and validated
+""")
 
 checked_kerberos = False
 kerberos = None
@@ -312,17 +353,60 @@
         self._wrap_socket_for_reporting(self.sock)
 
 
-# Build the appropriate socket wrapper for ssl
-try:
-    # python 2.6 introduced a better ssl package
-    import ssl
-    _ssl_wrap_socket = ssl.wrap_socket
-except ImportError:
-    # python versions prior to 2.6 don't have ssl and ssl.wrap_socket instead
-    # they use httplib.FakeSocket
-    def _ssl_wrap_socket(sock, key_file, cert_file):
-        ssl_sock = socket.ssl(sock, key_file, cert_file)
-        return httplib.FakeSocket(sock, ssl_sock)
+# These two methods were imported from Python 3.2's ssl module
+
+def _dnsname_to_pat(dn):
+    pats = []
+    for frag in dn.split(r'.'):
+        if frag == '*':
+            # When '*' is a fragment by itself, it matches a non-empty dotless
+            # fragment.
+            pats.append('[^.]+')
+        else:
+            # Otherwise, '*' matches any dotless fragment.
+            frag = re.escape(frag)
+            pats.append(frag.replace(r'\*', '[^.]*'))
+    return re.compile(r'\A' + r'\.'.join(pats) + r'\Z', re.IGNORECASE)
+
+
+def match_hostname(cert, hostname):
+    """Verify that *cert* (in decoded format as returned by
+    SSLSocket.getpeercert()) matches the *hostname*.  RFC 2818 rules
+    are mostly followed, but IP addresses are not accepted for *hostname*.
+
+    CertificateError is raised on failure. On success, the function
+    returns nothing.
+    """
+    if not cert:
+        raise ValueError("empty or no certificate")
+    dnsnames = []
+    san = cert.get('subjectAltName', ())
+    for key, value in san:
+        if key == 'DNS':
+            if _dnsname_to_pat(value).match(hostname):
+                return
+            dnsnames.append(value)
+    if not san:
+        # The subject is only checked when subjectAltName is empty
+        for sub in cert.get('subject', ()):
+            for key, value in sub:
+                # XXX according to RFC 2818, the most specific Common Name
+                # must be used.
+                if key == 'commonName':
+                    if _dnsname_to_pat(value).match(hostname):
+                        return
+                    dnsnames.append(value)
+    if len(dnsnames) > 1:
+        raise errors.CertificateError("hostname %r "
+            "doesn't match either of %s"
+            % (hostname, ', '.join(map(repr, dnsnames))))
+    elif len(dnsnames) == 1:
+        raise errors.CertificateError("hostname %r "
+            "doesn't match %r"
+            % (hostname, dnsnames[0]))
+    else:
+        raise errors.CertificateError("no appropriate commonName or "
+            "subjectAltName fields were found")
 
 
 class HTTPSConnection(AbstractHTTPConnection, httplib.HTTPSConnection):
@@ -345,7 +429,19 @@
             self.connect_to_origin()
 
     def connect_to_origin(self):
-        ssl_sock = _ssl_wrap_socket(self.sock, self.key_file, self.cert_file)
+        # FIXME JRV 2011-12-18: Use location config here?
+        config_stack = config.GlobalStack()
+        ca_certs = config_stack.get('ssl.ca_certs')
+        cert_reqs = config_stack.get('ssl.cert_reqs')
+        if cert_reqs == "none":
+            trace.warning("not checking SSL certificates for %s: %d",
+                self.host, self.port)
+        ssl_sock = ssl.wrap_socket(self.sock, self.key_file, self.cert_file,
+            cert_reqs=cert_reqs, ca_certs=ca_certs)
+        peer_cert = ssl_sock.getpeercert()
+        if cert_reqs == "required" or peer_cert:
+            match_hostname(peer_cert, self.host)
+
         # Wrap the ssl socket before anybody use it
         self._wrap_socket_for_reporting(ssl_sock)
 
 
=== modified file 'doc/en/release-notes/bzr-2.5.txt'
--- doc/en/release-notes/bzr-2.5.txt	2011-12-19 16:41:49 +0000
+++ doc/en/release-notes/bzr-2.5.txt	2011-12-20 09:40:32 +0000
@@ -408,6 +408,12 @@
   to 'line-based'. This replace the previous shelf UI only patch using
   ``INSIDE_EMACS``. (Benoît Pierre)
 
+* urllib-based HTTPS client connections now verify the server certificate
+  as well as the hostname. (Jelmer Vernooij)
+
+* urllib-based HTTPS connections are now the default.
+  (Jelmer Vernooij, #125055, #651161)
+
 Bug Fixes
 *********