tor/doc/tor-spec.txt
2003-04-18 18:57:22 +00:00

544 lines
23 KiB
Plaintext

$Id$
TOR (The Onion Router) Spec
Note: This is an attempt to specify TOR as it exists as implemented in
early March, 2003. It is not recommended that others implement this
design as it stands; future versions of TOR will implement improved
protocols.
0. Notation:
PK -- a public key.
SK -- a private key
K -- a key for a symmetric cypher
a|b -- concatenation of 'a' with 'b'.
a[i:j] -- Bytes 'i' through 'j'-1 (inclusive) of the string a.
All numeric values are encoded in network (big-endian) order.
Unless otherwise specified, all symmetric ciphers are DES in OFB
mode, with an IV of all 0 bytes. All asymmetric ciphers are RSA
with 1024-bit keys, and exponents of 65537.
[We will move to AES once we can assume everybody will have it. -RD]
1. System overview
[Something to start with here. Do feel free to change/expand. -RD]
Tor is an implementation of version 2 of Onion Routing.
Onion Routing is a connection-oriented anonymizing communication
service. Users build a layered block of asymmetric encryptions
(an "onion") which describes a source-routed path through a set of
nodes. Those nodes build a "virtual circuit" through the network, in which
each node knows its predecessor and successor, but no others. Traffic
flowing down the circuit is unwrapped by a symmetric key at each node,
which reveals the downstream node.
2. Connections
2.1. Establishing OR-to-OR connections
When one onion router opens a connection to another, the initiating
OR (called the 'client') and the listening OR (called the 'server')
perform the following handshake.
Before the handshake begins, the client and server know one
another's (1024-bit) public keys, IPV4 addresses, and ports.
1. Client connects to server:
The client generates a pair of 8-byte symmetric keys (one
[K_f] for the 'forward' stream from client to server, and one
[K_b] for the 'backward' stream from server to client.
The client then generates a 'Client authentication' message [M]
containing:
The client's published IPV4 address [4 bytes]
The client's published port [2 bytes]
The server's published IPV4 address [4 bytes]
The server's published port [2 bytes]
The forward key (K_f) [16 bytes]
The backward key (K_f) [16 bytes]
The maximum bandwidth (bytes/s) [4 bytes]
[Total: 48 bytes]
The client then RSA-encrypts the message with the server's
public key, and PKCS1 padding to given an encrypted message
[Commentary: 1024 bytes is probably too short, and this protocol can't
support IPv6. -NM]
[1024 is too short for a high-latency remailer; but perhaps it's
fine for us, given our need for speed and also given our greater
vulnerability to other attacks? Onions are infrequent enough now
that maybe we could handle it; but I worry it will impact
scalability, and handling more users is important.-RD]
The client then opens a TCP connection to the server, sends
the 128-byte RSA-encrypted data to the server, and waits for a
reply.
2. Server authenticates to client:
Upon receiving a TCP connection, the server waits to receive
128 bytes from the client. It decrypts the message with its
private key, and checks the PKCS1 padding. If the padding is
incorrect, or if the message's length is other than 32 bytes,
the server closes the TCP connection and stops handshaking.
The server then checks the list of known ORs for one with the
address and port given in the client's authentication. If no
such OR is known, or if the server is already connected to
that OR, the server closes the current TCP connection and
stops handshaking.
For later use, the server sets its keys for this connection,
setting K_f to the client's K_b, and K_b to the client's K_f.
The server then creates a server authentication message[M2] as
follows:
Modified client authentication [48 bytes]
A random nonce [N] [8 bytes]
[Total: 56 bytes]
The client authentication is generated from M by replacing
the client's preferred bandwidth [B_c] with the server's
preferred bandwidth [B_s], if B_s < B_c.
The server encrypts M2 with the client's public key (found
from the list of known routers), using PKCS1 padding.
The server sends the 128-byte encrypted message to the client,
and waits for a reply.
3. Client authenticates to server.
Once the client has received 128 bytes, it decrypts them with
its public key, and checks the PKCS1 padding. If the padding
is invalid, or the decrypted message's length is other than 40
bytes, the client closes the TCP connection.
The client checks that the addresses and keys in the reply
message are the same as the ones it originally sent. If not,
it closes the TCP connection.
The client updates the connection's bandwidth to that set by
the server, and generates the following authentication message [M3]:
The client's published IPV4 address [4 bytes]
The client's published port [2 bytes]
The server's published IPV4 address [4 bytes]
The server's published port [2 bytes]
The server-generated nonce [N] [8 bytes]
[Total: 20 bytes]
Once again, the client encrypts this message using the
server's public key and PKCS1 padding, and sends the resulting
128-byte message to the server.
4. Server checks client authentication
The server once again waits to receive 128 bytes from the
client, decrypts the message with its private key, and checks
the PKCS1 padding. If the padding is incorrect, or if the
message's length is other than 20 bytes, the server closes the
TCP connection and stops handshaking.
If the addresses in the decrypted message M3 match those in M
and M2, and if the nonce in M3 is the same as in M2, the
handshake is complete, and the client and server begin sending
cells to one another. Otherwise, the server closes the TCP
connection.
2.2. Establishing OP-to-OR connections
When an Onion Proxy (OP) needs to establish a connection to an OR,
the handshake is simpler because the OR does not need to verify the
OP's identity. The OP and OR establish the following steps:
1. OP connects to OR:
First, the OP generates a pair of 8-byte symmetric keys (one
[K_f] for the 'forward' stream from OP to OR, and one
[K_b] for the 'backward' stream from OR to OP).
The OP generates a message [M] in the following format:
Maximum bandwidth (bytes/s) [4 bytes]
Forward key [K_f] [16 bytes]
Backward key [K_b] [16 bytes]
[Total: 32 bytes]
The OP encrypts M with the OR's public key and PKCS1 padding,
opens a TCP connection to the OR's TCP port, and sends the
resulting 128-byte encrypted message to the OR.
2. OR receives keys:
When the OR receives a connection from an OP [This is on a
different port, right? How does it know the difference? -NM],
[Correct. The 'or_port' config variable specifies the OR port,
and the op_port variable specified the OP port. -RD]
it waits for 128 bytes of data, and decrypts the resulting
data with its private key, checking the PKCS1 padding. If the
padding is invalid, or the message is not 20 bytes long, the
OR closes the connection.
Otherwise, the connection is established, and the O is ready
to receive cells.
The server sets its keys for this connection, setting K_f to
the client's K_b, and K_b to the client's K_f.
2.3. Sending cells and link encryption
Once the handshake is complete, the ORs or OR and OP send cells
(specified below) to one another. Cells are sent serially,
encrypted with the 3DES-OFB keystream specified by the handshake
protocol. Over a connection, communicants encrypt outgoing cells
with the connection's K_f, and decrypt incoming cells with the
connection's K_b.
[Commentary: This means that OR/OP->OR connections are malleable; I
can flip bits in cells as they go across the wire, and see flipped
bits coming out the cells as they are decrypted at the next
server. I need to look more at the data format to see whether
this is exploitable, but if there's no integrity checking there
either, I suspect we may have an attack here. -NM]
[Yes, this protocol is open to tagging attacks. The payloads are
encrypted inside the network, so it's only at the edge node and beyond
that it's a worry. But adversaries can already count packets and
observe/modify timing. It's not worth putting in hashes; indeed, it
would be quite hard, because one of the sides of the circuit doesn't
know the keys that are used for de/encrypting at each hop, so couldn't
craft hashes anyway. See the Bandwidth Throttling (threat model)
thread on http://archives.seul.org/or/dev/Jul-2002/threads.html. -RD]
[Even if I don't control both sides of the connection, I can still
do evil stuff. For instance, if I can guess that a cell is a
TOPIC_COMMAND_BEGIN cell to www.slashdot.org:80 , I can change the
address and port to point to a machine I control. -NM]
3. Cell Packet format
The basic unit of communication for onion routers and onion
proxies is a fixed-width "cell". Each cell contains the following
fields:
ACI (anonymous circuit identifier) [2 bytes]
Command [1 byte]
Length [1 byte]
Sequence number (unused, set to 0) [4 bytes]
Payload (padded with 0 bytes) [248 bytes]
[Total size: 256 bytes]
The 'Command' field holds one of the following values:
0 -- PADDING (Padding) (See Sec 6.2)
1 -- CREATE (Create a circuit) (See Sec 4)
2 -- DATA (End-to-end data) (See Sec 5)
3 -- DESTROY (Stop using a circuit) (See Sec 4)
4 -- SENDME (For flow control) (See Sec 6.1)
The interpretation of 'Length' and 'Payload' depend on the type of
the cell.
PADDING: Length is 0; Payload is 248 bytes of 0's.
CREATE: Length is a value between 1 and 248; the first 'length'
bytes of payload contain a portion of an onion.
DATA: Length is a value between 4 and 248; the first 'length'
bytes of payload contain useful data.
DESTROY: Neither field is used.
SENDME: Length encodes a window size, payload is unused.
Unused fields are filled with 0 bytes. The payload is padded with
0 bytes.
PADDING cells are currently used to implement connection
keepalive. ORs and OPs send one another a PADDING cell every few
minutes.
CREATE and DESTROY cells are used to manage circuits; see section
4 below.
DATA cells are used to send commands and data along a circuit; see
section 5 below.
SENDME cells are used for flow control; see section 6 below.
4. Onions and circuit management
4.1. Setting up circuits
An onion is a multi-layered structure, with one layer for each node
in a circuit. Each (unencrypted) layer has the following fields:
Version [1 byte]
Port [2 bytes]
Address [4 bytes]
Expiration time [4 bytes]
Key seed material [16 bytes]
[Total: 27 bytes]
The value of Version is currently 2.
The port and address field denote the IPV4 address and port of
the next onion router in the circuit, or are set to 0 for the
last hop.
The expiration time is a number of seconds since the epoch (1
Jan 1970); by default, it is set to the current time plus one
day.
When constructing an onion to create a circuit from OR_1,
OR_2... OR_N, the onion creator performs the following steps:
1. Let M = 100 random bytes.
2. For I=N downto 1:
A. Create an onion layer L, setting Version=2,
ExpirationTime=now + 1 day, and Seed=16 random bytes.
If I=N, set Port=Address=0. Else, set Port and Address to
the IPV4 port and address of OR_{I+1}.
B. Let M = L | M.
C. Let K1_I = SHA1(Seed).
Let K2_I = SHA1(K1_I).
Let K3_I = SHA1(K2_I).
D. Encrypt the first 128 bytes of M with the RSA key of
OR_I, using no padding. Encrypt the remaining portion of
M with 3DES/OFB, using K1_I as a key and an all-0 IV.
3. M is now the onion.
To create a connection using the onion M, an OP or OR performs the
following steps:
1. If not already connected to the first router in the chain,
open a new connection to that router.
2. Choose an ACI not already in use on the connection with the
first router in the chain. If our address/port pair is
numerically higher than the address/port pair of the other
side, then let the high bit of the ACI be 1, else 0.
3. To send M over the wire, prepend a 4-byte integer containing
Len(M). Call the result M'. Let N=ceil(Len(M')/248).
Divide M' into N chunks, such that:
Chunk_I = M'[(I-1)*248:I*248] for 1 <= I <= N-1
Chunk_N = M'[(N-1)*248:Len(M')]
4. Send N CREATE cells along the connection, setting the ACI
on each to the selected ACI, setting the payload on each to
the corresponding 'Chunk_I', and setting the length on each
to the length of the payload.
Upon receiving a CREATE cell along a connection, an OR performs
the following steps:
1. If we already have an 'open' circuit along this connection
with this ACI, drop the cell.
Otherwise, if we have no circuit along this connection with
this ACI, let L = the integer value of the first 4 bytes of
the payload. Create a half-open circuit with this ACI, and
begin queueing CREATE cells for this circuit.
Otherwise, we have a half-open circuit. If the total payload
length of the CREATE cells for this circuit is exactly equal
to the onion length specified in the first cell (minus 4), then
process the onion. If it is more, then tear down the circuit.
2. Once we have a complete onion, decrypt the first 128 bytes
of the onion with this OR's RSA private key, and extract
the outmost onion layer. If the version, back cipher, or
forward cipher is unrecognized, or the expiration time is
in the past, then tear down the circuit (see section 4.2).
Compute K1 through K3 as above. Use K1 to decrypt the rest
of the onion using 3DES/OFB.
If we are not the exit node, remove the first layer from the
decrypted onion, and send the remainder to the next OR
on the circuit, as specified above. (Note that we'll
choose a different ACI for this circuit on the connection
with the next OR.)
As an optimization, OR implementations may delay processing onions
until a break in traffic allows time to do so without harming
network latency too greatly.
4.2. Tearing down circuits
Circuits are torn down when an unrecoverable error occurs along
the circuit, or when all topics on a circuit are closed and the
circuit's intended lifetime is over.
To tear down a circuit, an OR or OP sends a DESTROY cell with that
direction's ACI to the adjacent nodes on that circuit.
Upon receiving a DESTROY cell, an OR frees resources associated
with the corresponding circuit. If it's not the start or end of the
circuit, it sends a DESTROY cell for that circuit to the next OR in
the circuit. If the node is the start or end of the circuit, then
it tears down any associated edge connections (see section 5.1).
After a DESTROY cell has been processed, an OR ignores all data or
destroy cells for the corresponding circuit.
4.3. Routing data cells
When an OR receives a DATA cell, it checks the cell's ACI and
determines whether it has a corresponding circuit along that
connection. If not, the OR drops the DATA cell.
Otherwise, if the OR is not at the OP edge of the circuit (that is,
either an 'exit node' or a non-edge node), it de/encrypts the length
field and the payload with 3DES/OFB, as follows:
'Forward' data cell (same direction as onion):
Use K2 as key; encrypt.
'Back' data cell (opposite direction from onion):
Use K3 as key; decrypt.
Otherwise, if the data cell has arrived to the OP edge of the circuit,
the OP de/encrypts the length and payload fields with 3DES/OFB as
follows:
OP sends data cell:
For I=1...N, decrypt with K2_I.
OP receives data cell:
For I=N...1, encrypt with K3_I.
Edge nodes process the length and payload fields of DATA cells as
described in section 5 below.
5. Application connections and topic management
5.1. Topics and TCP streams
Within a circuit, the OP and the exit node use the contents of DATA
packets to tunnel TCP connections ("Topics") across circuits.
These connections are initiated by the OP.
The first 4 bytes of each data cell are reserved as follows:
Topic command [1 byte]
Unused, set to 0. [1 byte]
Topic ID [2 bytes]
The recognized topic commands are:
1 -- TOPIC_BEGIN
2 -- TOPIC_DATA
3 -- TOPIC_END
4 -- TOPIC_CONNECTED
5 -- TOPIC_SENDME
All DATA cells pertaining to the same tunneled connection have the
same topic ID.
To create a new anonymized TCP connection, the OP sends a
TOPIC_BEGIN data cell with a payload encoding the address and port
of the destination host. The payload format is:
ADDRESS | ':' | PORT | '\000'
where ADDRESS may be a DNS hostname, or an IPv4 address in
dotted-quad format; and where PORT is encoded in decimal.
Upon receiving this packet, the exit node resolves the address as
necessary, and opens a new TCP connection to the target port. If
the address cannot be resolved, or a connection can't be
established, the exit node replies with a TOPIC_END cell.
Otherwise, the exit node replies with a TOPIC_CONNECTED cell.
The OP waits for a TOPIC_CONNECTED cell before sending any data.
Once a connection has been established, the OP and exit node
package stream data in TOPIC_DATA cells, and upon receiving such
cells, echo their contents to the corresponding TCP stream.
[XXX Mention zlib encoding. -NM]
When one side of the TCP stream is closed, the corresponding edge
node sends a TOPIC_END cell along the circuit; upon receiving a
TOPIC_END cell, the edge node closes the corresponding TCP stream.
[This should probably become:
When one side of the TCP stream is closed, the corresponding edge
node sends a TOPIC_END cell along the circuit; upon receiving a
TOPIC_END cell, the edge node closes its side of the corresponding
TCP stream (by sending a FIN packet), but continues to accept and
package incoming data until both sides of the TCP stream are
closed. At that point, the edge node sends a second TOPIC_END
cell, and drops its record of the topic. -NM]
6. Flow control
6.1. Link throttling
As discussed above in section 2.1, ORs and OPs negotiate a maximum
bandwidth upon startup. The communicants only read up to that
number of bytes per second on average, though they may use mechanisms
to handle spikes (eg token buckets).
Communicants rely on TCP's default flow control to push back when they
stop reading, so nodes that don't obey this bandwidth limit can't do
too much damage.
6.2. Link padding
Currently nodes are not required to do any sort of link padding or
dummy traffic. Because strong attacks exist even with link padding,
and because link padding greatly increases the bandwidth requirements
for running a node, we plan to leave out link padding until this
tradeoff is better understood.
6.3. Circuit flow control
To control a circuit's bandwidth usage, each node keeps track of
how many data cells it is allowed to send to the next hop in the
circuit. This 'window' value is initially set to 1000 data cells
in each direction (cells that are not data cells do not affect
the window). Each edge node on a circuit sends a SENDME cell
(with length=100) every time it has received 100 data cells on the
circuit. When a node receives a SENDME cell for a circuit, it increases
the circuit's window in the corresponding direction (that is, for
sending data cells back in the direction from which the sendme arrived)
by the value of the cell's length field. If it's not an edge node,
it passes an equivalent SENDME cell to the next node in the circuit.
If the window value reaches 0 at the edge of a circuit, the OR stops
reading from the edge connections. (It may finish processing what
it's already read, and queue those cells for when a SENDME cell
arrives.) Otherwise (when not at the edge of a circuit), if the
window value is 0 and a data cell arrives, the node must tear down
the circuit.
6.4. Topic flow control
Edge nodes use TOPIC_SENDME data cells to implement end-to-end flow
control for individual connections across circuits. As with circuit
flow control, edge nodes begin with a window of cells (500) per
topic, and increment the window by a fixed value (50) upon receiving
a TOPIC_SENDME data cell. Edge nodes initiate TOPIC_SENDME data
cells when both a) the window is <= 450, and b) there are less than
ten cell payloads remaining to be flushed at that edge.
7. Directories and routers
7.1. Router descriptor format.
Line format : address ORPort OPPort APPort DirPort bandwidth(bytes/s)
followed by the router's public key.
ORport is where the router listens for other routers (speaking cells)
OPPort is where the router listens for onion proxies (speaking cells)
APPort is where the router listens for applications (speaking socks)
DirPort is where the router listens for directory download requests
Example:
moria.mit.edu 9001 9011 9021 9031 100000
-----BEGIN RSA PUBLIC KEY-----
MIGJAoGBAMBBuk1sYxEg5jLAJy86U3GGJ7EGMSV7yoA6mmcsEVU3pwTUrpbpCmwS
7BvovoY3z4zk63NZVBErgKQUDkn3pp8n83xZgEf4GI27gdWIIwaBjEimuJlEY+7K
nZ7kVMRoiXCbjL6VAtNa4Zy1Af/GOm0iCIDpholeujQ95xew7rQnAgMA//8=
-----END RSA PUBLIC KEY-----