tor/doc/tor-spec.txt

$Id$

                         Tor Protocol Specification

                              Roger Dingledine
                               Nick Mathewson

Note: This document aims to specify Tor as implemented in 0.1.2.1-alpha-dev
and later.  Future versions of Tor will implement improved protocols, and
compatibility is not guaranteed.

THIS DOCUMENT IS UNSTABLE.  Right now, we're revising the protocol to remove
a few long-standing limitations.  For the most stable current version of the
protocol, see tor-spec-v0.txt; current versions of Tor are backward-compatible.

This specification is not a design document; most design criteria
are not examined.  For more information on why Tor acts as it does,
see tor-design.pdf.

TODO for v1 revision:
      - Fix onionskin handshake scheme to be more mainstream, less nutty.
        Can we just do
            E(HMAC(g^x), g^x) rather than just E(g^x) ?
        No, that has the same flaws as before. We should send
            E(g^x, C) with random C and expect g^y, HMAC_C(K=g^xy).
        Better ask Ian; probably Stephen too.
      - Versioned CREATE and friends
      - Length on CREATE and friends
      - Versioning on circuits

TODO:
      - REASON_CONNECTFAILED should include an IP.
      - Copy prose from tor-design to make everything more readable.
      - Spec when we should rotate which keys (tls, link, etc)?

0. Preliminaries

0.1.  Notation and encoding

   PK -- a public key.
   SK -- a private key.
   K  -- a key for a symmetric cypher.

   a|b -- concatenation of 'a' and 'b'.

   [A0 B1 C2] -- a three-byte sequence, containing the bytes with
   hexadecimal values A0, B1, and C2, in that order.

   All numeric values are encoded in network (big-endian) order.

   H(m) -- a cryptographic hash of m.

0.2. Security parameters

   Tor uses a stream cipher, a public-key cipher, the Diffie-Hellman
   protocol, and a hash function.

   KEY_LEN -- the length of the stream cipher's key, in bytes.

   PK_ENC_LEN -- the length of a public-key encrypted message, in bytes.
   PK_PAD_LEN -- the number of bytes added in padding for public-key
     encryption, in bytes. (The largest number of bytes that can be encrypted
     in a single public-key operation is therefore PK_ENC_LEN-PK_PAD_LEN.)

   DH_LEN -- the number of bytes used to represent a member of the
     Diffie-Hellman group.
   DH_SEC_LEN -- the number of bytes used in a Diffie-Hellman private key (x).

   HASH_LEN -- the length of the hash function's output, in bytes.

   PAYLOAD_LEN -- The longest allowable cell payload, in bytes. (509)

   CELL_LEN -- The length of a Tor cell, in bytes.

0.3. Ciphers

   For a stream cipher, we use 128-bit AES in counter mode, with an IV of all
   0 bytes.

   For a public-key cipher, we use RSA with 1024-bit keys and a fixed
   exponent of 65537.  We use OAEP padding, with SHA-1 as its digest
   function.   (For OAEP padding, see
   ftp://ftp.rsasecurity.com/pub/pkcs/pkcs-1/pkcs-1v2-1.pdf)

   For Diffie-Hellman, we use a generator (g) of 2.  For the modulus (p), we
   use the 1024-bit safe prime from rfc2409 section 6.2 whose hex
   representation is:

     "FFFFFFFFFFFFFFFFC90FDAA22168C234C4C6628B80DC1CD129024E08"
     "8A67CC74020BBEA63B139B22514A08798E3404DDEF9519B3CD3A431B"
     "302B0A6DF25F14374FE1356D6D51C245E485B576625E7EC6F44C42E9"
     "A637ED6B0BFF5CB6F406B7EDEE386BFB5A899FA5AE9F24117C4B1FE6"
     "49286651ECE65381FFFFFFFFFFFFFFFF"

   As an optimization, implementations SHOULD choose DH private keys (x) of
   320 bits.  Implementations that do this MUST never use any DH key more
   than once.
   [May other implementations reuse their DH keys?? -RD]

   For a hash function, we use SHA-1.

   KEY_LEN=16.
   DH_LEN=128; DH_GROUP_LEN=40.
   PK_ENC_LEN=128; PK_PAD_LEN=42.
   HASH_LEN=20.

   When we refer to "the hash of a public key", we mean the SHA-1 hash of the
   DER encoding of an ASN.1 RSA public key (as specified in PKCS.1).

   All "random" values should be generated with a cryptographically strong
   random number generator, unless otherwise noted.

   The "hybrid encryption" of a byte sequence M with a public key PK is
   computed as follows:
      1. If M is less than PK_ENC_LEN-PK_PAD_LEN, pad and encrypt M with PK.
      2. Otherwise, generate a KEY_LEN byte random key K.
         Let M1 = the first PK_ENC_LEN-PK_PAD_LEN-KEY_LEN bytes of M,
         and let M2 = the rest of M.
         Pad and encrypt K|M1 with PK.  Encrypt M2 with our stream cipher,
         using the key K.  Concatenate these encrypted values.
   [XXX Note that this "hybrid encryption" approach does not prevent
   an attacker from adding or removing bytes to the end of M. It also
   allows attackers to modify the bytes not covered by the OAEP --
   see Goldberg's PET2006 paper for details. We will add a MAC to this
   scheme one day. -RD]

0.4. Other parameter values

   CELL_LEN=512

1. System overview

   Tor is a distributed overlay network designed to anonymize
   low-latency TCP-based applications such as web browsing, secure shell,
   and instant messaging. Clients choose a path through the network and
   build a ``circuit'', in which each node (or ``onion router'' or ``OR'')
   in the path knows its predecessor and successor, but no other nodes in
   the circuit.  Traffic flowing down the circuit is sent in fixed-size
   ``cells'', which are unwrapped by a symmetric key at each node (like
   the layers of an onion) and relayed downstream.

1.1. Protocol Versioning

   The node-to-node TLS-based "OR connection" protocol and the multi-hop
   "circuit" protocol are versioned quasi-independently.  (Certain versions
   of the circuit protocol may require a minimum version of the connection
   protocol to be used.)

   Version numbers are incremented for backward-incompatible protocol changes
   only.  Backward-compatible changes are generally implemented by adding
   additional fields to existing structures; implementations MUST ignore
   fields they do not expect.

   Parties negotiate OR connection versions as described below in sections
   4.1 and 4.2.

2. Connections

   Tor uses TLS for link encryption.  All implementations MUST support
   the TLS ciphersuite "TLS_EDH_RSA_WITH_DES_192_CBC3_SHA", and SHOULD
   support "TLS_DHE_RSA_WITH_AES_128_CBC_SHA" if it is available.
   Implementations MAY support other ciphersuites, but MUST NOT
   support any suite without ephemeral keys, symmetric keys of at
   least KEY_LEN bits, and digests of at least HASH_LEN bits.

   Even though the connection protocol is identical, we think of the
   initiator as either an onion router (OR) if it is willing to relay
   traffic for other Tor users, or an onion proxy (OP) if it only handles
   local requests. Onion proxies SHOULD NOT provide long-term-trackable
   identifiers in their handshakes.

   The connection initiator always sends a two-certificate chain,
   consisting of a
   certificate using a short-term connection key and a second, self-
   signed certificate containing the OR's identity key. The commonName of the
   first certificate is the OR's nickname, and the commonName of the second
   certificate is the OR's nickname, followed by a space and the string
   "<identity>".

   Implementations running 0.2.1.0-alpha-dev and earlier used an
   organizationName of "Tor" or "TOR".  Current implementations (which
   support the version negotiation protocol in section 4.1) MUST NOT
   have either of these values for their organizationName.

   All parties receiving certificates must confirm that the identity key is
   as expected.  (When initiating a connection, the expected identity key is
   the one given in the directory; when creating a connection because of an
   EXTEND cell, the expected identity key is the one given in the cell.)  If
   the key is not as expected, the party must close the connection.

   All parties SHOULD reject connections to or from ORs that have malformed
   or missing certificates.  ORs MAY accept or reject connections from OPs
   with malformed or missing certificates.

   Once a TLS connection is established, the two sides send cells
   (specified below) to one another.  Cells are sent serially.  All
   cells are CELL_LEN bytes long.  Cells may be sent embedded in TLS
   records of any size or divided across TLS records, but the framing
   of TLS records MUST NOT leak information about the type or contents
   of the cells.

   TLS connections are not permanent. Either side may close a connection
   if there are no circuits running over it and an amount of time
   (KeepalivePeriod, defaults to 5 minutes) has passed.

   (As an exception, directory servers may try to stay connected to all of
   the ORs -- though this will be phased out for the Tor 0.1.2.x release.)

3. Cell Packet format

   The basic unit of communication for onion routers and onion
   proxies is a fixed-width "cell".

   On a version 0 connection, each cell contains the following
   fields:

        CircID                                [2 bytes]
        Command                               [1 byte]
        Payload (padded with 0 bytes)         [PAYLOAD_LEN bytes]

   On a version 1 connection, each cell contains the following fields:

        CircID                                [3 bytes]
        Command                               [1 byte]
        Payload (padded with 0 bytes)         [PAYLOAD_LEN bytes]

   The CircID field determines which circuit, if any, the cell is
   associated with.

   The 'Command' field holds one of the following values:
         0 -- PADDING     (Padding)                 (See Sec 7.2)
         1 -- CREATE      (Create a circuit)        (See Sec 5.1)
         2 -- CREATED     (Acknowledge create)      (See Sec 5.1)
         3 -- RELAY       (End-to-end data)         (See Sec 5.5 and 6)
         4 -- DESTROY     (Stop using a circuit)    (See Sec 5.4)
         5 -- CREATE_FAST (Create a circuit, no PK) (See Sec 5.1)
         6 -- CREATED_FAST (Circuit created, no PK) (See Sec 5.1)
         7 -- VERSIONS    (Negotiate versions)      (See Sec 4.1)
         8 -- NETINFO     (Time and MITM-prevention) (See Sec 4.2)

   The interpretation of 'Payload' depends on the type of the cell.
      PADDING: Payload is unused.
      CREATE:  Payload contains the handshake challenge.
      CREATED: Payload contains the handshake response.
      RELAY:   Payload contains the relay header and relay body.
      DESTROY: Payload contains a reason for closing the circuit.
               (see 5.4)
   Upon receiving any other value for the command field, an OR must
   drop the cell.  [XXXX Versions prior to 0.1.0.?? logged a warning
   when dropping the cell; this is bad behavior. -NM]

   The payload is padded with 0 bytes.

   PADDING cells are currently used to implement connection keepalive.
   If there is no other traffic, ORs and OPs send one another a PADDING
   cell every few minutes.

   CREATE, CREATED, and DESTROY cells are used to manage circuits;
   see section 4 below.

   RELAY cells are used to send commands and data along a circuit; see
   section 5 below.

   VERSIONS cells are used to introduce parameters and characteristics of
   Tor clients and servers when connections are established.

4, Connection management

   Upon establishing a TLS connection, both parties immediately begin
   negotiating a connection protocol version and other connection parameters.

4.1. VERSIONS cells

   When a Tor connection is established, both parties normally send a
   VERSIONS cell before sending any other cells.  (But see below.)

         NumVersions            [1 byte]
         Versions               [NumVersions bytes]

   "Versions" is a sequence of NumVersions link connection protocol versions,
   each one byte long.  Parties should list all of the versions which they
   are able and willing to support.  Parties can only communicate if they
   have some connection protocol version in common.

   Version 0.1.x.y-alpha and earlier don't understand VERSIONS cells,
   and therefore don't support version negotiation.  Thus, waiting until
   the other side has sent a VERSIONS cell won't work for these servers:
   if they send no cells back, it is impossible to tell whether they
   have sent a VERSIONS cell that has been stalled, or whether they have
   dropped our own VERSIONS cell as unrecognized.  Thus, immediately after
   a TLS connection has been established, the parties check whether the
   other side has an obsolete certificate (organizationName equal to "Tor"
   or "TOR").  If the other party presented an obsolete certificate,
   we assume a v0 connection.  Otherwise, both parties send VERSIONS
   cells listing all their supported versions.  Upon receiving the
   other party's VERSIONS cell, the implementation begins using the
   highest-valued version common to both cells.  If the first cell from
   the other party is _not_ a VERSIONS cell, we assume a v0 protocol.

   Implementations MUST discard cells that are not the first cells sent on a
   connection.

4.2. MITM-prevention and time checking

   If we negotiate a v1 connection or higher, the first cell we send SHOULD
   be a NETINFO cell.  Implementations SHOULD NOT send NETINFO cells at other
   times.

   A NETINFO cell contains:
         Timestamp              [4 bytes]
         This OR's address      [variable]
         Other OR's address     [variable]

   Timestamp is the OR's current Unix time, in seconds since the epoch.  If
   an implementation receives time values from many validated ORs that
   indicate that its clock is skewed, it SHOULD try to warn the
   administrator.

   Each address contains Type/Length/Value as used in Section 6.4.  The first
   address is the address of the interface the party sending the VERSIONS cell
   used to connect to or accept connections from the other -- we include it
   to block a man-in-the-middle attack on TLS that lets an attacker bounce
   traffic through his own computers to enable timing and packet-counting
   attacks.

   The second address is the one that the party sending the VERSIONS cell
   believes the other has -- it can be used to learn what your IP address
   is if you have no other hints.

5. Circuit management

5.1. CREATE and CREATED cells

   Users set up circuits incrementally, one hop at a time. To create a
   new circuit, OPs send a CREATE cell to the first node, with the
   first half of the DH handshake; that node responds with a CREATED
   cell with the second half of the DH handshake plus the first 20 bytes
   of derivative key data (see section 5.2). To extend a circuit past
   the first hop, the OP sends an EXTEND relay cell (see section 5)
   which instructs the last node in the circuit to send a CREATE cell
   to extend the circuit.

   The payload for a CREATE cell is an 'onion skin', which consists
   of the first step of the DH handshake data (also known as g^x).
   This value is hybrid-encrypted (see 0.3) to Bob's public key, giving
   an onion-skin of:
       PK-encrypted:
         Padding padding               [PK_PAD_LEN bytes]
         Symmetric key                 [KEY_LEN bytes]
         First part of g^x             [PK_ENC_LEN-PK_PAD_LEN-KEY_LEN bytes]
       Symmetrically encrypted:
         Second part of g^x            [DH_LEN-(PK_ENC_LEN-PK_PAD_LEN-KEY_LEN)
                                           bytes]

   The relay payload for an EXTEND relay cell consists of:
         Address                       [4 bytes]
         Port                          [2 bytes]
         Onion skin                    [DH_LEN+KEY_LEN+PK_PAD_LEN bytes]
         Identity fingerprint          [HASH_LEN bytes]

   The port and address field denote the IPV4 address and port of the next
   onion router in the circuit; the public key hash is the hash of the PKCS#1
   ASN1 encoding of the next onion router's identity (signing) key.  (See 0.3
   above.)  (Including this hash allows the extending OR verify that it is
   indeed connected to the correct target OR, and prevents certain
   man-in-the-middle attacks.)

   The payload for a CREATED cell, or the relay payload for an
   EXTENDED cell, contains:
         DH data (g^y)                 [DH_LEN bytes]
         Derivative key data (KH)      [HASH_LEN bytes]   <see 5.2 below>

   The CircID for a CREATE cell is an arbitrarily chosen 2-byte integer,
   selected by the node (OP or OR) that sends the CREATE cell.  To prevent
   CircID collisions, when one OR sends a CREATE cell to another, it chooses
   from only one half of the possible values based on the ORs' public
   identity keys: if the sending OR has a lower key, it chooses a CircID with
   an MSB of 0; otherwise, it chooses a CircID with an MSB of 1.

   Public keys are compared numerically by modulus.

   As usual with DH, x and y MUST be generated randomly.

[
   To implement backward-compatible version negotiation, parties MUST
   drop CREATE cells with all-[00] onion-skins.
]

5.1.1. CREATE_FAST/CREATED_FAST cells

   When initializing the first hop of a circuit, the OP has already
   established the OR's identity and negotiated a secret key using TLS.
   Because of this, it is not always necessary for the OP to perform the
   public key operations to create a circuit.  In this case, the
   OP MAY send a CREATE_FAST cell instead of a CREATE cell for the first
   hop only.  The OR responds with a CREATED_FAST cell, and the circuit is
   created.

   A CREATE_FAST cell contains:

       Key material (X)    [HASH_LEN bytes]

   A CREATED_FAST cell contains:

       Key material (Y)    [HASH_LEN bytes]
       Derivative key data [HASH_LEN bytes] (See 5.2 below)

   The values of X and Y must be generated randomly.

   [Versions of Tor before 0.1.0.6-rc did not support these cell types;
    clients should not send CREATE_FAST cells to older Tor servers.]

   If an OR sees a circuit created with CREATE_FAST, the OR is sure to be the
   first hop of a circuit.  ORs SHOULD reject attempts to create streams with
   RELAY_BEGIN exiting the circuit at the first hop: letting Tor be used as a
   single hop proxy makes exit nodes a more attractive target for compromise.

5.2. Setting circuit keys

   Once the handshake between the OP and an OR is completed, both can
   now calculate g^xy with ordinary DH.  Before computing g^xy, both client
   and server MUST verify that the received g^x or g^y value is not degenerate;
   that is, it must be strictly greater than 1 and strictly less than p-1
   where p is the DH modulus.  Implementations MUST NOT complete a handshake
   with degenerate keys.  Implementations MUST NOT discard other "weak"
   g^x values.

   (Discarding degenerate keys is critical for security; if bad keys
   are not discarded, an attacker can substitute the server's CREATED
   cell's g^y with 0 or 1, thus creating a known g^xy and impersonating
   the server. Discarding other keys may allow attacks to learn bits of
   the private key.)

   (The mainline Tor implementation, in the 0.1.1.x-alpha series, discarded
   all g^x values less than 2^24, greater than p-2^24, or having more than
   1024-16 identical bits.  This served no useful purpose, and we stopped.)

   If CREATE or EXTEND is used to extend a circuit, the client and server
   base their key material on K0=g^xy, represented as a big-endian unsigned
   integer.

   If CREATE_FAST is used, the client and server base their key material on
   K0=X|Y.

   From the base key material K0, they compute KEY_LEN*2+HASH_LEN*3 bytes of
   derivative key data as
       K = H(K0 | [00]) | H(K0 | [01]) | H(K0 | [02]) | ...

   The first HASH_LEN bytes of K form KH; the next HASH_LEN form the forward
   digest Df; the next HASH_LEN 41-60 form the backward digest Db; the next
   KEY_LEN 61-76 form Kf, and the final KEY_LEN form Kb.  Excess bytes from K
   are discarded.

   KH is used in the handshake response to demonstrate knowledge of the
   computed shared key. Df is used to seed the integrity-checking hash
   for the stream of data going from the OP to the OR, and Db seeds the
   integrity-checking hash for the data stream from the OR to the OP. Kf
   is used to encrypt the stream of data going from the OP to the OR, and
   Kb is used to encrypt the stream of data going from the OR to the OP.

5.3. Creating circuits

   When creating a circuit through the network, the circuit creator
   (OP) performs the following steps:

      1. Choose an onion router as an exit node (R_N), such that the onion
         router's exit policy includes at least one pending stream that
         needs a circuit (if there are any).

      2. Choose a chain of (N-1) onion routers
         (R_1...R_N-1) to constitute the path, such that no router
         appears in the path twice.

      3. If not already connected to the first router in the chain,
         open a new connection to that router.

      4. Choose a circID not already in use on the connection with the
         first router in the chain; send a CREATE cell along the
         connection, to be received by the first onion router.

      5. Wait until a CREATED cell is received; finish the handshake
         and extract the forward key Kf_1 and the backward key Kb_1.

      6. For each subsequent onion router R (R_2 through R_N), extend
         the circuit to R.

   To extend the circuit by a single onion router R_M, the OP performs
   these steps:

      1. Create an onion skin, encrypted to R_M's public key.

      2. Send the onion skin in a relay EXTEND cell along
         the circuit (see section 5).

      3. When a relay EXTENDED cell is received, verify KH, and
         calculate the shared keys.  The circuit is now extended.

   When an onion router receives an EXTEND relay cell, it sends a CREATE
   cell to the next onion router, with the enclosed onion skin as its
   payload.  The initiating onion router chooses some circID not yet
   used on the connection between the two onion routers.  (But see
   section 5.1. above, concerning choosing circIDs based on
   lexicographic order of nicknames.)

   When an onion router receives a CREATE cell, if it already has a
   circuit on the given connection with the given circID, it drops the
   cell.  Otherwise, after receiving the CREATE cell, it completes the
   DH handshake, and replies with a CREATED cell.  Upon receiving a
   CREATED cell, an onion router packs it payload into an EXTENDED relay
   cell (see section 5), and sends that cell up the circuit.  Upon
   receiving the EXTENDED relay cell, the OP can retrieve g^y.

   (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.)

5.4. Tearing down circuits

   Circuits are torn down when an unrecoverable error occurs along
   the circuit, or when all streams on a circuit are closed and the
   circuit's intended lifetime is over.  Circuits may be torn down
   either completely or hop-by-hop.

   To tear down a circuit completely, an OR or OP sends a DESTROY
   cell to the adjacent nodes on that circuit, using the appropriate
   direction's circID.

   Upon receiving an outgoing DESTROY cell, an OR frees resources
   associated with the corresponding circuit. If it's not the end of
   the circuit, it sends a DESTROY cell for that circuit to the next OR
   in the circuit. If the node is the end of the circuit, then it tears
   down any associated edge connections (see section 6.1).

   After a DESTROY cell has been processed, an OR ignores all data or
   destroy cells for the corresponding circuit.

   To tear down part of a circuit, the OP may send a RELAY_TRUNCATE cell
   signaling a given OR (Stream ID zero).  That OR sends a DESTROY
   cell to the next node in the circuit, and replies to the OP with a
   RELAY_TRUNCATED cell.

   When an unrecoverable error occurs along one connection in a
   circuit, the nodes on either side of the connection should, if they
   are able, act as follows:  the node closer to the OP should send a
   RELAY_TRUNCATED cell towards the OP; the node farther from the OP
   should send a DESTROY cell down the circuit.

   The payload of a RELAY_TRUNCATED or DESTROY cell contains a single octet,
   describing why the circuit is being closed or truncated.  When sending a
   TRUNCATED or DESTROY cell because of another TRUNCATED or DESTROY cell,
   the error code should be propagated.  The origin of a circuit always sets
   this error code to 0, to avoid leaking its version.

   The error codes are:
     0 -- NONE            (No reason given.)
     1 -- PROTOCOL        (Tor protocol violation.)
     2 -- INTERNAL        (Internal error.)
     3 -- REQUESTED       (A client sent a TRUNCATE command.)
     4 -- HIBERNATING     (Not currently operating; trying to save bandwidth.)
     5 -- RESOURCELIMIT   (Out of memory, sockets, or circuit IDs.)
     6 -- CONNECTFAILED   (Unable to reach server.)
     7 -- OR_IDENTITY     (Connected to server, but its OR identity was not
                           as expected.)
     8 -- OR_CONN_CLOSED  (The OR connection that was carrying this circuit
                           died.)
     9 -- FINISHED        (The circuit has expired for being dirty or old.)
    10 -- TIMEOUT         (Circuit construction took too long)
    11 -- DESTROYED       (The circuit was destroyed w/o client TRUNCATE)
    12 -- NOSUCHSERVICE   (Request for unknown hidden service)

   [Versions of Tor prior to 0.1.0.11 didn't send reasons; implementations
   MUST accept empty TRUNCATED and DESTROY cells.]

5.5. Routing relay cells

   When an OR receives a RELAY cell, it checks the cell's circID and
   determines whether it has a corresponding circuit along that
   connection.  If not, the OR drops the RELAY 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 payload
   with the stream cipher, as follows:
        'Forward' relay cell (same direction as CREATE):
            Use Kf as key; decrypt.
        'Back' relay cell (opposite direction from CREATE):
            Use Kb as key; encrypt.
   Note that in counter mode, decrypt and encrypt are the same operation.

   The OR then decides whether it recognizes the relay cell, by
   inspecting the payload as described in section 6.1 below.  If the OR
   recognizes the cell, it processes the contents of the relay cell.
   Otherwise, it passes the decrypted relay cell along the circuit if
   the circuit continues.  If the OR at the end of the circuit
   encounters an unrecognized relay cell, an error has occurred: the OR
   sends a DESTROY cell to tear down the circuit.

   When a relay cell arrives at an OP, the OP decrypts the payload
   with the stream cipher as follows:
         OP receives data cell:
            For I=N...1,
                Decrypt with Kb_I.  If the payload is recognized (see
                section 6..1), then stop and process the payload.

   For more information, see section 6 below.

6. Application connections and stream management

6.1. Relay cells

   Within a circuit, the OP and the exit node use the contents of
   RELAY packets to tunnel end-to-end commands and TCP connections
   ("Streams") across circuits.  End-to-end commands can be initiated
   by either edge; streams are initiated by the OP.

   The payload of each unencrypted RELAY cell consists of:
         Relay command           [1 byte]
         'Recognized'            [2 bytes]
         StreamID                [2 bytes]
         Digest                  [4 bytes]
         Length                  [2 bytes]
         Data                    [CELL_LEN-14 bytes]

   The relay commands are:
         1 -- RELAY_BEGIN     [forward]
         2 -- RELAY_DATA      [forward or backward]
         3 -- RELAY_END       [forward or backward]
         4 -- RELAY_CONNECTED [backward]
         5 -- RELAY_SENDME    [forward or backward] [sometimes control]
         6 -- RELAY_EXTEND    [forward]             [control]
         7 -- RELAY_EXTENDED  [backward]            [control]
         8 -- RELAY_TRUNCATE  [forward]             [control]
         9 -- RELAY_TRUNCATED [backward]            [control]
        10 -- RELAY_DROP      [forward or backward] [control]
        11 -- RELAY_RESOLVE   [forward]
        12 -- RELAY_RESOLVED  [backward]
        13 -- RELAY_BEGIN_DIR [forward]

   Commands labelled as "forward" must only be sent by the originator
   of the circuit. Commands labelled as "backward" must only be sent by
   other nodes in the circuit back to the originator. Commands marked
   as either can be sent either by the originator or other nodes.

   The 'recognized' field in any unencrypted relay payload is always set
   to zero; the 'digest' field is computed as the first four bytes of
   the running digest of all the bytes that have been destined for
   this hop of the circuit or originated from this hop of the circuit,
   seeded from Df or Db respectively (obtained in section 5.2 above),
   and including this RELAY cell's entire payload (taken with the digest
   field set to zero).

   When the 'recognized' field of a RELAY cell is zero, and the digest
   is correct, the cell is considered "recognized" for the purposes of
   decryption (see section 5.5 above).

   (The digest does not include any bytes from relay cells that do
   not start or end at this hop of the circuit. That is, it does not
   include forwarded data. Therefore if 'recognized' is zero but the
   digest does not match, the running digest at that node should
   not be updated, and the cell should be forwarded on.)

   All RELAY cells pertaining to the same tunneled stream have the
   same stream ID.  StreamIDs are chosen arbitrarily by the OP.  RELAY
   cells that affect the entire circuit rather than a particular
   stream use a StreamID of zero -- they are marked in the table above
   as "[control]" style cells. (Sendme cells are marked as "sometimes
   control" because they can take include a StreamID or not depending
   on their purpose -- see Section 7.)

   The 'Length' field of a relay cell contains the number of bytes in
   the relay payload which contain real payload data. The remainder of
   the payload is padded with NUL bytes.

   If the RELAY cell is recognized but the relay command is not
   understood, the cell must be dropped and ignored. Its contents
   still count with respect to the digests, though. [Before
   0.1.1.10, Tor closed circuits when it received an unknown relay
   command. Perhaps this will be more forward-compatible. -RD]

6.2. Opening streams and transferring data

   To open a new anonymized TCP connection, the OP chooses an open
   circuit to an exit that may be able to connect to the destination
   address, selects an arbitrary StreamID not yet used on that circuit,
   and constructs a RELAY_BEGIN cell with a payload encoding the address
   and port of the destination host.  The payload format is:

         ADDRESS | ':' | PORT | [00]

   where  ADDRESS can be a DNS hostname, or an IPv4 address in
   dotted-quad format, or an IPv6 address surrounded by square brackets;
   and where PORT is encoded in decimal.

   [What is the [00] for? -NM]
   [It's so the payload is easy to parse out with string funcs -RD]

   Upon receiving this cell, 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 RELAY_END cell.  (See 6.4 below.)
   Otherwise, the exit node replies with a RELAY_CONNECTED cell, whose
   payload is in one of the following formats:
       The IPv4 address to which the connection was made [4 octets]
       A number of seconds (TTL) for which the address may be cached [4 octets]
    or
       Four zero-valued octets [4 octets]
       An address type (6)     [1 octet]
       The IPv6 address to which the connection was made [16 octets]
       A number of seconds (TTL) for which the address may be cached [4 octets]
   [XXXX Versions of Tor before 0.1.1.6 ignore and do not generate the TTL
   field.  No version of Tor currently generates the IPv6 format.

   Tor servers before 0.1.2.0 set the TTL field to a fixed value.  Later
   versions set the TTL to the last value seen from a DNS server, and expire
   their own cached entries after a fixed interval.  This prevents certain
   attacks.]

   The OP waits for a RELAY_CONNECTED cell before sending any data.
   Once a connection has been established, the OP and exit node
   package stream data in RELAY_DATA cells, and upon receiving such
   cells, echo their contents to the corresponding TCP stream.
   RELAY_DATA cells sent to unrecognized streams are dropped.

   Relay RELAY_DROP cells are long-range dummies; upon receiving such
   a cell, the OR or OP must drop it.

6.2.1. Opening a directory stream

   If a Tor server is a directory server, it should respond to a
   RELAY_BEGIN_DIR cell as if it had received a BEGIN cell requesting a
   connection to its directory port.  RELAY_BEGIN_DIR cells ignore exit
   policy, since the stream is local to the Tor process.

   If the Tor server is not running a directory service, it should respond
   with a REASON_NOTDIRECTORY RELAY_END cell.

   Clients MUST generate an all-zero payload for RELAY_BEGIN_DIR cells,
   and servers MUST ignore the payload.

   [RELAY_BEGIN_DIR was not supported before Tor 0.1.2.2-alpha; clients
   SHOULD NOT send it to routers running earlier versions of Tor.]

6.3. Closing streams

   When an anonymized TCP connection is closed, or an edge node
   encounters error on any stream, it sends a 'RELAY_END' cell along the
   circuit (if possible) and closes the TCP connection immediately.  If
   an edge node receives a 'RELAY_END' cell for any stream, it closes
   the TCP connection completely, and sends nothing more along the
   circuit for that stream.

   The payload of a RELAY_END cell begins with a single 'reason' byte to
   describe why the stream is closing, plus optional data (depending on
   the reason.)  The values are:

       1 -- REASON_MISC           (catch-all for unlisted reasons)
       2 -- REASON_RESOLVEFAILED  (couldn't look up hostname)
       3 -- REASON_CONNECTREFUSED (remote host refused connection) [*]
       4 -- REASON_EXITPOLICY     (OR refuses to connect to host or port)
       5 -- REASON_DESTROY        (Circuit is being destroyed)
       6 -- REASON_DONE           (Anonymized TCP connection was closed)
       7 -- REASON_TIMEOUT        (Connection timed out, or OR timed out
                                   while connecting)
       8 -- (unallocated) [**]
       9 -- REASON_HIBERNATING    (OR is temporarily hibernating)
      10 -- REASON_INTERNAL       (Internal error at the OR)
      11 -- REASON_RESOURCELIMIT  (OR has no resources to fulfill request)
      12 -- REASON_CONNRESET      (Connection was unexpectedly reset)
      13 -- REASON_TORPROTOCOL    (Sent when closing connection because of
                                   Tor protocol violations.)
      14 -- REASON_NOTDIRECTORY   (Client sent RELAY_BEGIN_DIR to a
                                   non-directory server.)

   (With REASON_EXITPOLICY, the 4-byte IPv4 address or 16-byte IPv6 address
   forms the optional data; no other reason currently has extra data.
   As of 0.1.1.6, the body also contains a 4-byte TTL.)

   OPs and ORs MUST accept reasons not on the above list, since future
   versions of Tor may provide more fine-grained reasons.

   [*] Older versions of Tor also send this reason when connections are
       reset.
   [**] Due to a bug in versions of Tor through 0095, error reason 8 must
        remain allocated until that version is obsolete.

   --- [The rest of this section describes unimplemented functionality.]

   Because TCP connections can be half-open, we follow an equivalent
   to TCP's FIN/FIN-ACK/ACK protocol to close streams.

   An exit connection can have a TCP stream in one of three states:
   'OPEN', 'DONE_PACKAGING', and 'DONE_DELIVERING'.  For the purposes
   of modeling transitions, we treat 'CLOSED' as a fourth state,
   although connections in this state are not, in fact, tracked by the
   onion router.

   A stream begins in the 'OPEN' state.  Upon receiving a 'FIN' from
   the corresponding TCP connection, the edge node sends a 'RELAY_FIN'
   cell along the circuit and changes its state to 'DONE_PACKAGING'.
   Upon receiving a 'RELAY_FIN' cell, an edge node sends a 'FIN' to
   the corresponding TCP connection (e.g., by calling
   shutdown(SHUT_WR)) and changing its state to 'DONE_DELIVERING'.

   When a stream in already in 'DONE_DELIVERING' receives a 'FIN', it
   also sends a 'RELAY_FIN' along the circuit, and changes its state
   to 'CLOSED'.  When a stream already in 'DONE_PACKAGING' receives a
   'RELAY_FIN' cell, it sends a 'FIN' and changes its state to
   'CLOSED'.

   If an edge node encounters an error on any stream, it sends a
   'RELAY_END' cell (if possible) and closes the stream immediately.

6.4. Remote hostname lookup

   To find the address associated with a hostname, the OP sends a
   RELAY_RESOLVE cell containing the hostname to be resolved.  (For a reverse
   lookup, the OP sends a RELAY_RESOLVE cell containing an in-addr.arpa
   address.)  The OR replies with a RELAY_RESOLVED cell containing a status
   byte, and any number of answers.  Each answer is of the form:
       Type   (1 octet)
       Length (1 octet)
       Value  (variable-width)
       TTL    (4 octets)
   "Length" is the length of the Value field.
   "Type" is one of:
      0x00 -- Hostname
      0x04 -- IPv4 address
      0x06 -- IPv6 address
      0xF0 -- Error, transient
      0xF1 -- Error, nontransient

    If any answer has a type of 'Error', then no other answer may be given.

    The RELAY_RESOLVE cell must use a nonzero, distinct streamID; the
    corresponding RELAY_RESOLVED cell must use the same streamID.  No stream
    is actually created by the OR when resolving the name.

7. Flow control

7.1. Link throttling

   Each node should do appropriate bandwidth throttling to keep its
   user happy.

   Communicants rely on TCP's default flow control to push back when they
   stop reading.

7.2. Link padding

   Link padding can be created by sending PADDING cells along the
   connection; relay cells of type "DROP" can be used for long-range
   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.

7.3. Circuit-level flow control

   To control a circuit's bandwidth usage, each OR keeps track of
   two 'windows', consisting of how many RELAY_DATA cells it is
   allowed to package for transmission, and how many RELAY_DATA cells
   it is willing to deliver to streams outside the network.
   Each 'window' value is initially set to 1000 data cells
   in each direction (cells that are not data cells do not affect
   the window).  When an OR is willing to deliver more cells, it sends a
   RELAY_SENDME cell towards the OP, with Stream ID zero.  When an OR
   receives a RELAY_SENDME cell with stream ID zero, it increments its
   packaging window.

   Each of these cells increments the corresponding window by 100.

   The OP behaves identically, except that it must track a packaging
   window and a delivery window for every OR in the circuit.

   An OR or OP sends cells to increment its delivery window when the
   corresponding window value falls under some threshold (900).

   If a packaging window reaches 0, the OR or OP stops reading from
   TCP connections for all streams on the corresponding circuit, and
   sends no more RELAY_DATA cells until receiving a RELAY_SENDME cell.
[this stuff is badly worded; copy in the tor-design section -RD]

7.4. Stream-level flow control

   Edge nodes use RELAY_SENDME cells to implement end-to-end flow
   control for individual connections across circuits. Similarly to
   circuit-level flow control, edge nodes begin with a window of cells
   (500) per stream, and increment the window by a fixed value (50)
   upon receiving a RELAY_SENDME cell. Edge nodes initiate RELAY_SENDME
   cells when both a) the window is <= 450, and b) there are less than
   ten cell payloads remaining to be flushed at that edge.


A.1. Differences between spec and implementation

- The current specification requires all ORs to have IPv4 addresses, but
  allows servers to exit and resolve to IPv6 addresses, and to declare IPv6
  addresses in their exit policies.  The current codebase has no IPv6
  support at all.

B. Things that should change in a later version of the Tor protocol

B.1. ... but which will require backward-incompatible change

  - Circuit IDs should be longer.
  - IPv6 everywhere.
  - Maybe, keys should be longer.
    - Maybe, key-length should be adjustable.  How to do this without
      making anonymity suck?
  - Drop backward compatibility.
  - We should use a 128-bit subgroup of our DH prime.
  - Handshake should use HMAC.
  - Multiple cell lengths.
  - Ability to split circuits across paths (If this is useful.)
  - SENDME windows should be dynamic.

  - Directory
     - Stop ever mentioning socks ports

B.1. ... and that will require no changes

   - Mention multiple addr/port combos
   - Advertised outbound IP?
   - Migrate streams across circuits.

B.2. ... and that we have no idea how to do.

   - UDP (as transport)
   - UDP (as content)
   - Use a better AES mode that has built-in integrity checking,
     doesn't grow with the number of hops, is not patented, and
     is implemented and maintained by smart people.