gvpe/doc/gvpe.protocol.7.pod

=head1 The GNU-VPE Protocols

=head1 Overview

GVPE can make use of a number of protocols. One of them is the GNU VPE
protocol which is used to authenticate tunnels and send encrypted data
packets. This protocol is described in more detail the second part of this
document.

The first part of this document describes the transport protocols which
are used by GVPE to send it's data packets over the network.

=head1 PART 1: Transport protocols

GVPE offers a wide range of transport protocols that can be used to
interchange data between nodes. Protocols differ in their overhead, speed,
reliability, and robustness.

The following sections describe each transport protocol in more
detail. They are sorted by overhead/efficiency, the most efficient
transport is listed first:

=head2 RAW IP

This protocol is the best choice, performance-wise, as the minimum
overhead per packet is only 38 bytes.

It works by sending the VPN payload using raw IP frames (using the
protocol set by C<ip-proto>).

Using raw IP frames has the drawback that many firewalls block "unknown"
protocols, so this transport only works if you have full IP connectivity
between nodes.

=head2 ICMP

This protocol offers very low overhead (minimum 42 bytes), and can
sometimes tunnel through firewalls when other protocols can not.

It works by prepending an ICMP header with type C<icmp-type> and a code
of C<255>. The default C<icmp-type> is C<echo-reply>, so the resulting
packets look like echo replies, which looks rather strange to network
administrators.

This transport should only be used if other transports (i.e. raw IP) are
not available or undesirable (due to their overhead).

=head2 UDP

This is a good general choice for the transport protocol as UDP packets
tunnel well through most firewalls and routers, and the overhead per
packet is moderate (minimum 58 bytes).

It should be used if RAW IP is not available.

=head2 TCP

This protocol is a very bad choice, as it not only has high overhead (more
than 60 bytes), but the transport also retries on it's own, which leads
to congestion when the link has moderate packet loss (as both the TCP
transport and the tunneled traffic will retry, increasing congestion more
and more). It also has high latency and is quite inefficient.

It's only useful when tunneling through firewalls that block better
protocols. If a node doesn't have direct internet access but a HTTP proxy
that supports the CONNECT method it can be used to tunnel through a web
proxy. For this to work, the C<tcp-port> should be C<443> (C<https>), as
most proxies do not allow connections to other ports.

It is an abuse of the usage a proxy was designed for, so make sure you are
allowed to use it for GVPE.

This protocol also has server and client sides. If the C<tcp-port> is
set to zero, other nodes cannot connect to this node directly. If the
C<tcp-port> is non-zero, the node can act both as a client as well as a
server.

=head2 DNS

B<WARNING:> Parsing and generating DNS packets is rather tricky. The code
almost certainly contains buffer overflows and other, likely exploitable,
bugs. You have been warned.

This is the worst choice of transport protocol with respect to overhead
(overhead can be 2-3 times higher than the transferred data), and latency
(which can be many seconds). Some DNS servers might not be prepared to
handle the traffic and drop or corrupt packets. The client also has to
constantly poll the server for data, so the client will constantly create
traffic even if it doesn't need to transport packets.

In addition, the same problems as the TCP transport also plague this
protocol.

Its only use is to tunnel through firewalls that do not allow direct
internet access. Similar to using a HTTP proxy (as the TCP transport
does), it uses a local DNS server/forwarder (given by the C<dns-forw-host>
configuration value) as a proxy to send and receive data as a client,
and an C<NS> record pointing to the GVPE server (as given by the
C<dns-hostname> directive).

The only good side of this protocol is that it can tunnel through most
firewalls mostly undetected, iff the local DNS server/forwarder is sane
(which is true for most routers, wireless LAN gateways and nameservers).

Fine-tuning needs to be done by editing C<src/vpn_dns.C> directly.

=head1 PART 2: The GNU VPE protocol

This section, unfortunately, is not yet finished, although the protocol
is stable (until bugs in the cryptography are found, which will likely
completely change the following description). Nevertheless, it should give
you some overview over the protocol.

=head2 Anatomy of a VPN packet

The exact layout and field lengths of a VPN packet is determined at
compile time and doesn't change. The same structure is used for all
transport protocols, be it RAWIP or TCP.

 +------+------+--------+------+
 | HMAC | TYPE | SRCDST | DATA |
 +------+------+--------+------+

The HMAC field is present in all packets, even if not used (e.g. in auth
request packets), in which case it is set to all zeroes. The checksum
itself is calculated over the TYPE, SRCDST and DATA fields in all cases.

The TYPE field is a single byte and determines the purpose of the packet
(e.g. RESET, COMPRESSED/UNCOMPRESSED DATA, PING, AUTH REQUEST/RESPONSE,
CONNECT REQUEST/INFO etc.).

SRCDST is a three byte field which contains the source and destination
node IDs (12 bits each).

The DATA portion differs between each packet type, naturally, and is the
only part that can be encrypted. Data packets contain more fields, as
shown:

 +------+------+--------+------+-------+------+
 | HMAC | TYPE | SRCDST | RAND | SEQNO | DATA |
 +------+------+--------+------+-------+------+

RAND is a sequence of fully random bytes, used to increase the entropy of
the data for encryption purposes.

SEQNO is a 32-bit sequence number. It is negotiated at every connection
initialization and starts at some random 31 bit value. GVPE currently uses
a sliding window of 512 packets/sequence numbers to detect reordering,
duplication and replay attacks.

The encryption is done on RAND+SEQNO+DATA in CBC mode with zero IV (or,
equivalently, the IV is RAND+SEQNO, encrypted with the block cipher,
unless RAND size is decreased or increased over the default value).

The random prefix itself is generated by using AES in CTR mode with a
random key and starting value, which should make them unpredictable even
before encrypting them again. The sequence number additionally ensures
that the IV is unique.

=head2 The authentication/key exchange protocol

Before nodes can exchange packets, they need to establish authenticity of
the other side and a key. Every node has a private RSA key and the public
RSA keys of all other nodes.

When a node wants to establish a connection to another node, it sends an
RSA-OEAP-encrypted challenge and an ECDH key. The other node replies with
it's own ECDH key and a HKDF of the challange and both ECDH keys to proof
it's identity.

The remote node enganges in exactly the same protocol. When both nodes
have exchanged their challenge and verified the response, they calculate a
cipher key and a HMAC key and start exchanging data packets.

In detail, the challenge consist of:

  RSA-OAEP (SEQNO MAC CIPHER SALT EXTRA-AUTH) ECDH1

That is, it encrypts (with the public key of the remote node) an initial
sequence number for data packets, key material for the HMAC key, key
material for the cipher key, a salt used by the HKDF (as shown later) and
some extra random bytes that are unused except for authentication. It also
sends the public key of a curve25519 exchange.

The remote node decrypts the RSA data, generates it's own ECDH key (ECDH2), and
replies with:

  HKDF-Expand (HKDF-Extract (ECDH2, RSA), ECDH1, AUTH_DIGEST_SIZE) ECDH2

That is, it extracts from the decrypted RSA challenge, using it's ECDH
key as salt, and then expands using the requesting node's ECDH1 key. The
resulting has is returned as a proof that the node could decrypt the RSA
challenge data, together with the ECDH key.

After both nodes have done this to each other, they calculate the shared
ECDH secrets, cipher and HMAC keys for the session (each
node generates two cipher and HMAC keys, one for sending and one for
receiving).

The HMAC key for sending is generated as follow:

   HMAC_KEY = HKDF-Expand (HKDF-Extract (REMOTE_SALT, MAC ECDH_SECRET), info, HMAC_MD_SIZE)

It extracts from MAC and ECDH_SECRET using the I<remote> SALT, then
expands using a static info string.

The cipher key is generated in the same way, except using the CIPHER part
of the original challenge.

The result of this process is to authenticate each node to the other
node, while exchanging keys using both RSA and ECDH, the latter providing
perfect forward secrecy.

The protocol has been overdesigned where this was possible without
increasing implementation complexity, in an attempt to protect against
implementation or protocol failures. For example, if the ECDH challenge
was found to be flawed, perfect forward secrecy would be lost, but
the data would still be protected. Likewise, standard algorithms and
implementations are used where possible.

=head2 Retrying

When there is no response to an auth request, the node will send auth
requests in bursts with an exponential back-off. After some time it will
resort to PING packets, which are very small (8 bytes + protocol header)
and lightweight (no RSA operations required). A node that receives ping
requests from an unconnected peer will respond by trying to create a
connection.

In addition to the exponential back-off, there is a global rate-limit on
a per-IP base. It allows long bursts but will limit total packet rate to
something like one control packet every ten seconds, to avoid accidental
floods due to protocol problems (like a RSA key file mismatch between two
nodes).

The intervals between retries are limited by the C<max-retry>
configuration value. A node with C<connect> = C<always> will always retry,
a node with C<connect> = C<ondemand> will only try (and re-try) to connect
as long as there are packets in the queue, usually this limits the retry
period to C<max-ttl> seconds.

Sending packets over the VPN will reset the retry intervals as well, which
means as long as somebody is trying to send packets to a given node, GVPE
will try to connect every few seconds.

=head2 Routing and Protocol translation

The GVPE routing algorithm is easy: there isn't much routing to speak
of: When routing packets to another node, GVPE tries the following
options, in order:

=over 4

=item If the two nodes should be able to reach each other directly (common
protocol, port known), then GVPE will send the packet directly to the
other node.

=item If this isn't possible (e.g. because the node doesn't have a
C<hostname> or known port), but the nodes speak a common protocol and a
router is available, then GVPE will ask a router to "mediate" between both
nodes (see below).

=item If a direct connection isn't possible (no common protocols) or
forbidden (C<deny-direct>) and there are any routers, then GVPE will try
to send packets to the router with the highest priority that is connected
already I<and> is able (as specified by the config file) to connect
directly to the target node.

=item If no such router exists, then GVPE will simply send the packet to
the node with the highest priority available.

=item Failing all that, the packet will be dropped.

=back

A host can usually declare itself unreachable directly by setting it's
port number(s) to zero. It can declare other hosts as unreachable by using
a config-file that disables all protocols for these other hosts. Another
option is to disable all protocols on that host in the other config files.

If two hosts cannot connect to each other because their IP address(es)
are not known (such as dial-up hosts), one side will send a I<mediated>
connection request to a router (routers must be configured to act as
routers!), which will send both the originating and the destination host
a connection info request with protocol information and IP address of the
other host (if known). Both hosts will then try to establish a direct
connection to the other peer, which is usually possible even when both
hosts are behind a NAT gateway.

Routing via other nodes works because the SRCDST field is not encrypted,
so the router can just forward the packet to the destination host. Since
each host uses it's own private key, the router will not be able to
decrypt or encrypt packets, it will just act as a simple router and
protocol translator.


Revision:	1.12
Committed:	Fri Jul 19 21:13:44 2013 UTC (10 years, 10 months ago) by root
Branch:	MAIN
Changes since 1.11:	+13 -1 lines
Log Message:	* empty log message *
#	Content
1	=head1 The GNU-VPE Protocols
2
3	=head1 Overview
4
5	GVPE can make use of a number of protocols. One of them is the GNU VPE
6	protocol which is used to authenticate tunnels and send encrypted data
7	packets. This protocol is described in more detail the second part of this
8	document.
9
10	The first part of this document describes the transport protocols which
11	are used by GVPE to send it's data packets over the network.
12
13	=head1 PART 1: Transport protocols
14
15	GVPE offers a wide range of transport protocols that can be used to
16	interchange data between nodes. Protocols differ in their overhead, speed,
17	reliability, and robustness.
18
19	The following sections describe each transport protocol in more
20	detail. They are sorted by overhead/efficiency, the most efficient
21	transport is listed first:
22
23	=head2 RAW IP
24
25	This protocol is the best choice, performance-wise, as the minimum
26	overhead per packet is only 38 bytes.
27
28	It works by sending the VPN payload using raw IP frames (using the
29	protocol set by C<ip-proto>).
30
31	Using raw IP frames has the drawback that many firewalls block "unknown"
32	protocols, so this transport only works if you have full IP connectivity
33	between nodes.
34
35	=head2 ICMP
36
37	This protocol offers very low overhead (minimum 42 bytes), and can
38	sometimes tunnel through firewalls when other protocols can not.
39
40	It works by prepending an ICMP header with type C<icmp-type> and a code
41	of C<255>. The default C<icmp-type> is C<echo-reply>, so the resulting
42	packets look like echo replies, which looks rather strange to network
43	administrators.
44
45	This transport should only be used if other transports (i.e. raw IP) are
46	not available or undesirable (due to their overhead).
47
48	=head2 UDP
49
50	This is a good general choice for the transport protocol as UDP packets
51	tunnel well through most firewalls and routers, and the overhead per
52	packet is moderate (minimum 58 bytes).
53
54	It should be used if RAW IP is not available.
55
56	=head2 TCP
57
58	This protocol is a very bad choice, as it not only has high overhead (more
59	than 60 bytes), but the transport also retries on it's own, which leads
60	to congestion when the link has moderate packet loss (as both the TCP
61	transport and the tunneled traffic will retry, increasing congestion more
62	and more). It also has high latency and is quite inefficient.
63
64	It's only useful when tunneling through firewalls that block better
65	protocols. If a node doesn't have direct internet access but a HTTP proxy
66	that supports the CONNECT method it can be used to tunnel through a web
67	proxy. For this to work, the C<tcp-port> should be C<443> (C<https>), as
68	most proxies do not allow connections to other ports.
69
70	It is an abuse of the usage a proxy was designed for, so make sure you are
71	allowed to use it for GVPE.
72
73	This protocol also has server and client sides. If the C<tcp-port> is
74	set to zero, other nodes cannot connect to this node directly. If the
75	C<tcp-port> is non-zero, the node can act both as a client as well as a
76	server.
77
78	=head2 DNS
79
80	B<WARNING:> Parsing and generating DNS packets is rather tricky. The code
81	almost certainly contains buffer overflows and other, likely exploitable,
82	bugs. You have been warned.
83
84	This is the worst choice of transport protocol with respect to overhead
85	(overhead can be 2-3 times higher than the transferred data), and latency
86	(which can be many seconds). Some DNS servers might not be prepared to
87	handle the traffic and drop or corrupt packets. The client also has to
88	constantly poll the server for data, so the client will constantly create
89	traffic even if it doesn't need to transport packets.
90
91	In addition, the same problems as the TCP transport also plague this
92	protocol.
93
94	Its only use is to tunnel through firewalls that do not allow direct
95	internet access. Similar to using a HTTP proxy (as the TCP transport
96	does), it uses a local DNS server/forwarder (given by the C<dns-forw-host>
97	configuration value) as a proxy to send and receive data as a client,
98	and an C<NS> record pointing to the GVPE server (as given by the
99	C<dns-hostname> directive).
100
101	The only good side of this protocol is that it can tunnel through most
102	firewalls mostly undetected, iff the local DNS server/forwarder is sane
103	(which is true for most routers, wireless LAN gateways and nameservers).
104
105	Fine-tuning needs to be done by editing C<src/vpn_dns.C> directly.
106
107	=head1 PART 2: The GNU VPE protocol
108
109	This section, unfortunately, is not yet finished, although the protocol
110	is stable (until bugs in the cryptography are found, which will likely
111	completely change the following description). Nevertheless, it should give
112	you some overview over the protocol.
113
114	=head2 Anatomy of a VPN packet
115
116	The exact layout and field lengths of a VPN packet is determined at
117	compile time and doesn't change. The same structure is used for all
118	transport protocols, be it RAWIP or TCP.
119
120	+------+------+--------+------+
121	\| HMAC \| TYPE \| SRCDST \| DATA \|
122	+------+------+--------+------+
123
124	The HMAC field is present in all packets, even if not used (e.g. in auth
125	request packets), in which case it is set to all zeroes. The checksum
126	itself is calculated over the TYPE, SRCDST and DATA fields in all cases.
127
128	The TYPE field is a single byte and determines the purpose of the packet
129	(e.g. RESET, COMPRESSED/UNCOMPRESSED DATA, PING, AUTH REQUEST/RESPONSE,
130	CONNECT REQUEST/INFO etc.).
131
132	SRCDST is a three byte field which contains the source and destination
133	node IDs (12 bits each).
134
135	The DATA portion differs between each packet type, naturally, and is the
136	only part that can be encrypted. Data packets contain more fields, as
137	shown:
138
139	+------+------+--------+------+-------+------+
140	\| HMAC \| TYPE \| SRCDST \| RAND \| SEQNO \| DATA \|
141	+------+------+--------+------+-------+------+
142
143	RAND is a sequence of fully random bytes, used to increase the entropy of
144	the data for encryption purposes.
145
146	SEQNO is a 32-bit sequence number. It is negotiated at every connection
147	initialization and starts at some random 31 bit value. GVPE currently uses
148	a sliding window of 512 packets/sequence numbers to detect reordering,
149	duplication and replay attacks.
150
151	The encryption is done on RAND+SEQNO+DATA in CBC mode with zero IV (or,
152	equivalently, the IV is RAND+SEQNO, encrypted with the block cipher,
153	unless RAND size is decreased or increased over the default value).
154
155	The random prefix itself is generated by using AES in CTR mode with a
156	random key and starting value, which should make them unpredictable even
157	before encrypting them again. The sequence number additionally ensures
158	that the IV is unique.
159
160	=head2 The authentication/key exchange protocol
161
162	Before nodes can exchange packets, they need to establish authenticity of
163	the other side and a key. Every node has a private RSA key and the public
164	RSA keys of all other nodes.
165
166	When a node wants to establish a connection to another node, it sends an
167	RSA-OEAP-encrypted challenge and an ECDH key. The other node replies with
168	it's own ECDH key and a HKDF of the challange and both ECDH keys to proof
169	it's identity.
170
171	The remote node enganges in exactly the same protocol. When both nodes
172	have exchanged their challenge and verified the response, they calculate a
173	cipher key and a HMAC key and start exchanging data packets.
174
175	In detail, the challenge consist of:
176
177	RSA-OAEP (SEQNO MAC CIPHER SALT EXTRA-AUTH) ECDH1
178
179	That is, it encrypts (with the public key of the remote node) an initial
180	sequence number for data packets, key material for the HMAC key, key
181	material for the cipher key, a salt used by the HKDF (as shown later) and
182	some extra random bytes that are unused except for authentication. It also
183	sends the public key of a curve25519 exchange.
184
185	The remote node decrypts the RSA data, generates it's own ECDH key (ECDH2), and
186	replies with:
187
188	HKDF-Expand (HKDF-Extract (ECDH2, RSA), ECDH1, AUTH_DIGEST_SIZE) ECDH2
189
190	That is, it extracts from the decrypted RSA challenge, using it's ECDH
191	key as salt, and then expands using the requesting node's ECDH1 key. The
192	resulting has is returned as a proof that the node could decrypt the RSA
193	challenge data, together with the ECDH key.
194
195	After both nodes have done this to each other, they calculate the shared
196	ECDH secrets, cipher and HMAC keys for the session (each
197	node generates two cipher and HMAC keys, one for sending and one for
198	receiving).
199
200	The HMAC key for sending is generated as follow:
201
202	HMAC_KEY = HKDF-Expand (HKDF-Extract (REMOTE_SALT, MAC ECDH_SECRET), info, HMAC_MD_SIZE)
203
204	It extracts from MAC and ECDH_SECRET using the I<remote> SALT, then
205	expands using a static info string.
206
207	The cipher key is generated in the same way, except using the CIPHER part
208	of the original challenge.
209
210	The result of this process is to authenticate each node to the other
211	node, while exchanging keys using both RSA and ECDH, the latter providing
212	perfect forward secrecy.
213
214	The protocol has been overdesigned where this was possible without
215	increasing implementation complexity, in an attempt to protect against
216	implementation or protocol failures. For example, if the ECDH challenge
217	was found to be flawed, perfect forward secrecy would be lost, but
218	the data would still be protected. Likewise, standard algorithms and
219	implementations are used where possible.
220
221	=head2 Retrying
222
223	When there is no response to an auth request, the node will send auth
224	requests in bursts with an exponential back-off. After some time it will
225	resort to PING packets, which are very small (8 bytes + protocol header)
226	and lightweight (no RSA operations required). A node that receives ping
227	requests from an unconnected peer will respond by trying to create a
228	connection.
229
230	In addition to the exponential back-off, there is a global rate-limit on
231	a per-IP base. It allows long bursts but will limit total packet rate to
232	something like one control packet every ten seconds, to avoid accidental
233	floods due to protocol problems (like a RSA key file mismatch between two
234	nodes).
235
236	The intervals between retries are limited by the C<max-retry>
237	configuration value. A node with C<connect> = C<always> will always retry,
238	a node with C<connect> = C<ondemand> will only try (and re-try) to connect
239	as long as there are packets in the queue, usually this limits the retry
240	period to C<max-ttl> seconds.
241
242	Sending packets over the VPN will reset the retry intervals as well, which
243	means as long as somebody is trying to send packets to a given node, GVPE
244	will try to connect every few seconds.
245
246	=head2 Routing and Protocol translation
247
248	The GVPE routing algorithm is easy: there isn't much routing to speak
249	of: When routing packets to another node, GVPE tries the following
250	options, in order:
251
252	=over 4
253
254	=item If the two nodes should be able to reach each other directly (common
255	protocol, port known), then GVPE will send the packet directly to the
256	other node.
257
258	=item If this isn't possible (e.g. because the node doesn't have a
259	C<hostname> or known port), but the nodes speak a common protocol and a
260	router is available, then GVPE will ask a router to "mediate" between both
261	nodes (see below).
262
263	=item If a direct connection isn't possible (no common protocols) or
264	forbidden (C<deny-direct>) and there are any routers, then GVPE will try
265	to send packets to the router with the highest priority that is connected
266	already I<and> is able (as specified by the config file) to connect
267	directly to the target node.
268
269	=item If no such router exists, then GVPE will simply send the packet to
270	the node with the highest priority available.
271
272	=item Failing all that, the packet will be dropped.
273
274	=back
275
276	A host can usually declare itself unreachable directly by setting it's
277	port number(s) to zero. It can declare other hosts as unreachable by using
278	a config-file that disables all protocols for these other hosts. Another
279	option is to disable all protocols on that host in the other config files.
280
281	If two hosts cannot connect to each other because their IP address(es)
282	are not known (such as dial-up hosts), one side will send a I<mediated>
283	connection request to a router (routers must be configured to act as
284	routers!), which will send both the originating and the destination host
285	a connection info request with protocol information and IP address of the
286	other host (if known). Both hosts will then try to establish a direct
287	connection to the other peer, which is usually possible even when both
288	hosts are behind a NAT gateway.
289
290	Routing via other nodes works because the SRCDST field is not encrypted,
291	so the router can just forward the packet to the destination host. Since
292	each host uses it's own private key, the router will not be able to
293	decrypt or encrypt packets, it will just act as a simple router and
294	protocol translator.
295
296