Don't unregister p11-kit PIN callback until vpninfo is finished with

Unregistering in openconnect_close_https() meant that when we reconnect to
the server, we lose the PIN callback. And then when we connect again, if
GnuTLS is asking us for the PIN on every attempt to touch the key, we fail
because there's no PIN handler.

So add a 'final' flag to openconnect_close_https(). Use this to clean up
library.c::openconnect_close_https() a little, too.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Import updated gnutls_pkcs12_simple_parse() from GnuTLS

Changes corresponding to commit 6c82bf34 in GnuTLS master, imported with
permission from Nikos to use under LGPLv2.1.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use X-DTLS-MTU response from server as well as X-CSTP-MTU

Currently we take a very naïve approach: we just use the higher of the
two. Normally the DTLS MTU will be larger. Theoretically, perhaps we
ought to actually change the MTU of the interface according to whether
DTLS is currently connected or not? That seems cumbersome, and is almost
impossible if we aren't running as root.

So what *should* we do with packets which are "too big" for the CSTP
MTU, if they arrive while DTLS is down? Drop them? And try to fake an
ICMP "too big" or "fragmentation needed" response? Fragment them? Please
$DEITY no. The sanest thing to do would seem to be just to send them
down the CSTP link even though they'll end up fragmented into more than
one TCP packet.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Update changelog

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add $CISCO_SPLIT_DNS environment variable for vpnc-script

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add gnutls.c and openssl.c to EXTRA_DIST too

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add gnutls_pkcs12 to dist

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Make 'make dist' work for out-of-tree build

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Set X-CSTP-Base-MTU: for new servers

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Don't hard-code cipher type in GnuTLS DTLS

Add an array with the two cipher labels (AES128-SHA and DES-CBC3-SHA) that
I've been able to test. The server doesn't seem to accept anything else
that we ask for.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix git-tree deps for version.c in out-of-tree build

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix generation of version.c for out-of-tree builds

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Change Requires: to Requires.private: in openconnect.pc

There's no need for users of the library to directly link with anything else
that we use.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix cbdata argument to process_auth_form()

I just introduced lots of bugs... oops.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Kill old openconnect_vpninfo_new()

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Bump library soname to libopenconnect.so.2

With this, the certificates are now an opaque type and callers are not
permitted to access them directly. Take the opportunity to also rename
openconnect_init_openssl() to openconnect_init_ssl().

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove last traces of special UI and PIN handling from main.c

Absolutely everything should now be proxied onto the ->process_auth_form()
function, so there's no need to handle anything directly.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use p11-kit for directing PIN request to process_auth_form()

Set a 'pin-source' attribute which identifies the vpninfo structure, and
register a handler which converts it to an auth form for the GUI to process.

If the URI we are given already contains a pin_source then theoretically
we don't override it; we assume the caller knew what they were doing. In
practice, p11_kit_get_pin_source() seems to be returning NULL even when
the attribute *is* set, so we always override it.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix GnuTLS PKCS#11 PIN request function

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS now uses gnutls_session_set_premaster()

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Show correct path to vpnc-script in the man page

Insert the actual path to vpnc-script that is compiled into the
openconnect executable.

Signed-off-by: Mike Miller <mtmiller@ieee.org>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add UI handling for OpenSSL TPM keys

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Allow '--with-gnutls' in configure

No need to require '--with-gnutls=shibboleet' any more; we have some
confidence that the GnuTLS support is actually working so we can let
non-hackers discover it.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add DTLS support for GnuTLS

This requires the patches I just sent to Nikos...

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Split out OpenSSL_specific start_dtls_handshake() function

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Check for gnutls_pkcs12_simple_parse() in GnuTLS

Our modifications made it upstream...

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use request_passphrase() for OpenSSL PEM files

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix config fetch

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix config_arg handling

The ->cert_password field must always be allocated, and it turns out I never
did fix the keep_config_arg() macro to do the right thing for options from
a file, despite deliberately introducing it for precisely that purpose.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use request_passphrase() for OpenSSL PKCS#12

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Move request_passphrase() to ssl.c

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Update comment about gnutls_x509_privkey_import_pkcs8() password handling

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix FreeBSD tun handling with net.link.tun.devfs_cloning=0

Try to use SIOCIFCREATE to create an interface if it doesn't already exists.
Also try opening /dev/tun to get the next available device, before falling
back to the loop over tun0-tun255.

There is still strangeness here; sometimes the interface doesn't get an
IPv6 link-local address, and the IFDISABLED flag remains set.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Allow interface name to be specified on *BSD

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix FreeBSD compile

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

OpenSSL: Don't include root CA in the supporting evidence; only intermediates

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS: Don't include root CA in the supporting evidence; only intermediates

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS: Split logging of additional certs into a separate loop

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add completely untested PIN callback for GnuTLS

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Stop using OpenSSL UI for user interaction

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

openconnect_set_xmlsha1() takes a const char *

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix GnuTLS request_passphrase() if no UI callback function

If it's NULL, don't call it. Also change 'gnutls' to 'ssl' in the auth_id,
since we may end up using this on the OpenSSL side too.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Reduce priority of GnuTLS certificate verify failure message to PRG_INFO

We don't want to see it in the auth-dialog UI; it's handled explicitly.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Bump API version, advertise get_cert_DER() and get_cert_details() functions

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix GnuTLS select() during handshake

It was using vpninfo->ssl_fd even though that's not set yet.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove stray debug printf

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix build on systems without O_CLOEXEC

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Move openconnect_SSL_printf() to ssl.c

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix vpninfo->peer_cert handling for GnuTLS

Stash the peer cert in verify_peer() so we can refer to it later.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove duplicate socket connect code from gnutls.c

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add openssl.c and gnutls.c to POTFILES

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add GnuTLS to changelog

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS: Import more than one certificate from PEM file

If the PEM file has extra "supporting" CAs, then import those and use them
too.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS: Load trusted CAs before loading certificate

We'll need them present when we load the certificate, because that's when
we search through them for supporting certs.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove unused workaround_openssl_certchain_bug() function

For GnuTLS, this is done as we load the certificate.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

GnuTLS: Add supporting certificates from PKCS#12 file

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Make GnuTLS parse_pkcs12() return extra certificates from the PKCS#12 too

Create a separate list, return them for the caller to do with as it sees fit.

This also cleans up the error handling a little. When this was a purely
internal GnuTLS function, it was fine to leave things (like *key) allocated
and return an error. If my intention is to make this exportable, then it
ought to clean up after itself when returning an error.

I think this actually fixes a potential memory leak for the GnuTLS internal
caller of this function, too.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Import pkcs12_parse() function from GnuTLS to fix PKCS#12 handling

An immediate effect is that this fixes the checking of cert expiry for
PKCS#12 certificates.

But it also means we can include the full supporting chain of
intermediate CAs (which has to be pre-assembled before we ever call
gnutls_certificate_set_x509_key() and can't be appended later), and we
can use the extra certs from the PKCS#12 file too, which parse_pkcs12()
currently doesn't bother to give us.

The plan is to fix parse_pkcs12(), submit the changes back upstream and
make it an exported function there, then stick a version-conditional on
our local copy and look forward to the day when we can rip it out again.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add server certificate validation for GnuTLS

It's broken with trust chains at the moment, at least with GnuTLS
2.12.x, because it looks up issuer certs by *name* and then when it
picks the wrong one the signature unsurprisingly fails. And then it
returns GNUTLS_CERT_INVALID without any specific *reason* for the
failure, which is even more joyful. At least with OpenSSL I can get a
reason string out of it.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Clean up GnuTLS PKCS#12 handling a little

Also try other types if gnutls_pkcs12_verify_mac() returns anything other
than GNUTLS_E_MAC_VERIFY_FAILED.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Check cert expiry, at least for PEM certs

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add client certificate support for GnuTLS

Argh. Why is there not just a function I can call to do this all *for* me?
249 lines of code for this one, which is more than the OpenSSL one I ranted
about at http://www.advogato.org/person/dwmw2/diary/205.html

Oh well, at least the password handling is *slightly* more consistent, if
not entirely so.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Improve GnuTLS compatibility options

TLSv1.0, no safe renegotiation, no padding.

For some reason, large amounts of padding are causing the Intel servers to
kick me off — although gnutls-cli is allowed to use large amounts of padding
with getting disconnected, and I can't see *why* there's a difference.

So there's something else odd going on here, and disabling padding is just
a workaround. I bet I forget about this, and I bet it comes back to bite
me one day. And it'll serve me right for being lazy and not following it
up properly right now. But still, there's plenty more GnuTLS porting work
to be done and I've spent long enough staring at packet traces already
today.

Disable safe renegotiation because we've previously observed that some
servers are behind crappy firewalls that'll block *any* extension.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Make CSTP connection in a single SSL record

By creating a buffer with the request and sending it in a single SSL record,
I roughly halve the amount of time it takes for the round trip from 215ms
to 116ms.

Introduce a buf_append() function to help with processing the buffer, since
I shouldn't just be using sprintf() like other places do. Will fix those
next...

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Refuse to build with GnuTLS < 2.12.16

We need the fix for gnutls_record_get_direction() or we end up sitting in
select() waiting for a read, when the blockage was actually on a *write*.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add barely functional GnuTLS support

It has no DTLS, doesn't do any server certificate validation, doesn't
support client certificates and there are odd bugs with it even in the
bits that *are* implemented. But we're getting there...

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Introduce semi-opaque OPENCONNECT_X509 type in library API

We offer functions to do everything that a user might want to do with the
cert, including one that returns it in DER form. The *only* reason this
isn't a completely opaque type is backward-compatibility.

When we change the soname, it'll be opaque. For now, let it actually be
an X509* for OpenSSL or a gnutls_x509_crt_t for GnuTLS.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add SSL library definition to CFLAGS in openconnect.pc

If openconnect.h is going to reference this, it needs to be set reliably.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove OpenSSL dependency from http.c

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Disable DTLS for GnuTLS build for now

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Move OpenSSL-specific functions from ssl.c to openssl.c

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Split out connect_https_socket() function from openconnect_open_https()

This can be used by the GnuTLS version too.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add openconnect_random() function

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add openconnect_get_cert_DER() function

This translates a cert into an SSL-library-agnostic form, so that the caller
can then process it using their own choice of tools.

As with the new openconnect_get_cert_details(), this isn't marked as a
public function yet because we anticipate more changes to the API.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Be more self-sufficient with header inclusions

Don't rely on things that are implicitly included through OpenSSL headers.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Move basic process_auth_form() out to main.c

There's no need for it to be in the library, and it uses OpenSSL UI.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add openconnect_sha1() function and use it instead of using OpenSSL directly

This also adds openssl.c that OpenSSL-specific functions will migrate to.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add openconnect_get_cert_details() function

Another aspect of the certificate handling becomes ssl-library-agnostic.

This is marked OPENCONNECT_PRIVATE for now. It probably *won't* be private,
but there are other changes to come and probably an soname bump, so there's
no point in exporting it for now.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Move peer_cert handling to openconnect_open_https()

There's no real need to do this in openconnect_obtain_cookie(). It doesn't
really matter if we do it for other connections, since any connections we
make *after* obtaining the cookie will be to the same server anyway.

This moves another OpenSSL-specific snippet out of what should be generic
code.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix API major/minor handling for out-of-source-tree build

We need to look in ${srcdir}/openconnect.h, not just openconnect.h

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add gnutls support to build system

Don't get excited; this is *only* in the build system. It won't build at all.
But we have to start somewhere.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use openconnect_open_https() and openconnect_close_https() better.

Use them unconditionally, without checking ->https_ssl first, and use them
in some places instead of open-coding the same thing.

This makes the code slightly more agnostic to the choice of SSL library.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix non-blocking support in fetch_config()

Rarely likely to matter, but writing the GET request for the config wasn't
coping with -EAGAIN.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Make openconnect_open_https() and openconnect_close_https() more forgiving.

If openconnect_open_https() is called with the connection already open,
return immediate success. Thus, the caller doesn't have to poke at
vpninfo->https_ssl to check it.

And if openconnect_close_https() is called when the connection isn't open,
don't attempt to close/free things that don't exist.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove libopenconnect.map from EXTRA_DIST

Signed-off-by: Mike Miller <mtmiller@ieee.org>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Add --with-pkgconfigdir to allow different environments like FreeBSD

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Tag version 3.20

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix invalid progress callback on socket cancellation

Oops! This caused the KDE auth-dialog to die because we pass it back
the wrong pointer.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Stash peer certificate before fetching HTTP response

If the server closes the connection by giving an HTTP 1.0-style response,
then the SSL connection will be gone by the time the GUI auth dialog calls
openconnect_get_peer_cert(). So remember it in order to give it out later.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Tag version 3.19

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Update changelog

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Update translations from Transifex

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Use SOCK_CLOEXEC when opening TCP socket

We were already setting FD_CLOEXEC but we should use SOCK_CLOEXEC where
it's available, so there's no race condition in a threaded environment.

When SOCK_CLOEXEC isn't available, set FD_CLOEXEC as soon as possible
after the socket() call to at least reduce the race window.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix --non-inter option so it still allows username/password from command line

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Fix error message when too many command line arguments

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Update translations from Transifex

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Call BIO_set_nbio() for SSL BIO at startup

Oops, this was still in CSTP code where we *used* to make the socket
non-blocking, and now it should be like that from the beginning.

It's not entirely clear what difference it makes; all my testing of the
non-blocking code through authentication and CSTP connection was working
fine in non-blocking mode without it, when all I'd done was set O_NONBLOCK
on the socket.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Clean up BIO_set_nbio() calls for DTLS

It's not necessary to do it twice for the same BIO, and it's not necessary
to get that BIO back from the SSL with SSL_get_[rw]bio() when we already
have it in a local variable.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Namespace cleanup: s/set_http_proxy/openconnect_set_http_proxy/

Now the library namespace is entirely prefixed by openconnect_ with no
inappropriate pollution. At least on platforms where the linker version
scripts work.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

Remove internal_parse_url() from the library exports.

The only thing that main.c was really doing differently to the public
openconnect_parse_url() function was allowing 'urlpath' to be superseded
by the --usergroup command line argument. Which we can handle simply
by storing that in a separate variable and applying it afterwards.

The other thing it did differently was check that the scheme is https.
But openconnect_parse_url() arguably should have been doing that anyway.

Fix potential memory leak of old strings in openconnect_parse_url()
while we're at it.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>