tor-parallel-relay-conn/doc/HACKING

0. Intro.
Onion Routing is still very much in development stages. This document
aims to get you started in the right direction if you want to understand
the code, add features, fix bugs, etc.

Read the README file first, so you can get familiar with the basics.

The pieces.

  Routers. Onion routers, as far as the 'tor' program is concerned,
  are a bunch of data items that are loaded into the router_array when
  the program starts. Periodically it downloads a new set of routers
  from a directory server, and updates the router_array. When a new OR
  connection is started (see below), the relevant information is copied
  from the router struct to the connection struct.

  Connections. A connection is a long-standing tcp socket between
  nodes. A connection is named based on what it's connected to -- an "OR
  connection" has an onion router on the other end, an "OP connection" has
  an onion proxy on the other end, an "exit connection" has a website or
  other server on the other end, and an "AP connection" has an application
  proxy (and thus a user) on the other end.

  Circuits. A circuit is a path over the onion routing
  network. Applications can connect to one end of the circuit, and can
  create exit connections at the other end of the circuit. AP and exit
  connections have only one circuit associated with them (and thus these
  connection types are closed when the circuit is closed), whereas OP and
  OR connections multiplex many circuits at once, and stay standing even
  when there are no circuits running over them.

  Streams. Streams are specific conversations between an AP and an exit.
  Streams are multiplexed over circuits.

  Cells. Some connections, specifically OR and OP connections, speak
  "cells". This means that data over that connection is bundled into 256
  byte packets (8 bytes of header and 248 bytes of payload). Each cell has
  a type, or "command", which indicates what it's for.

Robustness features.

[XXX no longer up to date]
 Bandwidth throttling. Each cell-speaking connection has a maximum
  bandwidth it can use, as specified in the routers.or file. Bandwidth
  throttling can occur on both the sender side and the receiving side. If
  the LinkPadding option is on, the sending side sends cells at regularly
  spaced intervals (e.g., a connection with a bandwidth of 25600B/s would
  queue a cell every 10ms). The receiving side protects against misbehaving
  servers that send cells more frequently, by using a simple token bucket:

  Each connection has a token bucket with a specified capacity. Tokens are
  added to the bucket each second (when the bucket is full, new tokens
  are discarded.) Each token represents permission to receive one byte
  from the network --- to receive a byte, the connection must remove a
  token from the bucket. Thus if the bucket is empty, that connection must
  wait until more tokens arrive. The number of tokens we add enforces a
  longterm average rate of incoming bytes, yet we still permit short-term
  bursts above the allowed bandwidth. Currently bucket sizes are set to
  ten seconds worth of traffic.

  The bandwidth throttling uses TCP to push back when we stop reading.
  We extend it with token buckets to allow more flexibility for traffic
  bursts.

 Data congestion control. Even with the above bandwidth throttling,
  we still need to worry about congestion, either accidental or intentional.
  If a lot of people make circuits into same node, and they all come out
  through the same connection, then that connection may become saturated
  (be unable to send out data cells as quickly as it wants to). An adversary
  can make a 'put' request through the onion routing network to a webserver
  he owns, and then refuse to read any of the bytes at the webserver end
  of the circuit. These bottlenecks can propagate back through the entire
  network, mucking up everything.

  (See the tor-spec.txt document for details of how congestion control
  works.)

  In practice, all the nodes in the circuit maintain a receive window
  close to maximum except the exit node, which stays around 0, periodically
  receiving a sendme and reading more data cells from the webserver.
  In this way we can use pretty much all of the available bandwidth for
  data, but gracefully back off when faced with multiple circuits (a new
  sendme arrives only after some cells have traversed the entire network),
  stalled network connections, or attacks.

  We don't need to reimplement full tcp windows, with sequence numbers,
  the ability to drop cells when we're full etc, because the tcp streams
  already guarantee in-order delivery of each cell. Rather than trying
  to build some sort of tcp-on-tcp scheme, we implement this minimal data
  congestion control; so far it's enough.

 Router twins. In many cases when we ask for a router with a given
  address and port, we really mean a router who knows a given key. Router
  twins are two or more routers that share the same private key. We thus
  give routers extra flexibility in choosing the next hop in the circuit: if
  some of the twins are down or slow, it can choose the more available ones.

  Currently the code tries for the primary router first, and if it's down,
  chooses the first available twin.

Coding conventions:

 Log convention: use only these four log severities.

  ERR is if something fatal just happened.
  WARNING is something bad happened, but we're still running. The
    bad thing is either a bug in the code, an attack or buggy
    protocol/implementation of the remote peer, etc. The operator should
    examine the bad thing and try to correct it.
  (No error or warning messages should be expected. I expect most people
    to run on -l warning eventually. If a library function is currently
    called such that failure always means ERR, then the library function
    should log WARNING and let the caller log ERR.)
  INFO means something happened (maybe bad, maybe ok), but there's nothing
    you need to (or can) do about it.
  DEBUG is for everything louder than INFO.