Data flow in the Tor process

We read bytes from the network, we write bytes to the network. For the most part, the bytes we write correspond roughly to bytes we have read, with bits of cryptography added in.

The rest is a matter of details.

Connections and buffers: reading, writing, and interpreting.

At a low level, Tor's networking code is based on "connections". Each connection represents an object that can send or receive network-like events. For the most part, each connection has a single underlying TCP stream (I'll discuss counterexamples below).

A connection that behaves like a TCP stream has an input buffer and an output buffer. Incoming data is written into the input buffer ("inbuf"); data to be written to the network is queued on an output buffer ("outbuf").

Buffers are implemented in buffers.c. Each of these buffers is implemented as a linked queue of memory extents, in the style of classic BSD mbufs, or Linux skbufs.

A connection's reading and writing can be enabled or disabled. Under the hood, this functionality is implemented using libevent events: one for reading, one for writing. These events are turned on/off in main.c, in the functions connection_{start,stop}_{reading,writing}.

When a read or write event is turned on, the main libevent loop polls the kernel, asking which sockets are ready to read or write. (This polling happens in the event_base_loop() call in run_main_loop_once() in main.c.) When libevent finds a socket that's ready to read or write, it invokes conn_{read,write}_callback(), also in main.c

These callback functions delegate to connection_handle_read() and connection_handle_write() in connection.c, which read or write on the network as appropriate, possibly delegating to openssl.

After data is read or written, or other event occurs, these connection_handle_read_write() functions call logic functions whose job is to respond to the information. Some examples included:

These functions then call into specific implementations depending on the type of the connection. For example, if the connection is an edge_connection_t, connection_reached_eof() will call connection_edge_reached_eof().

Note: "Also there are bufferevents!" We have vestigial code for an alternative low-level networking implementation, based on Libevent's evbuffer and bufferevent code. These two object types take on (most of) the roles of buffers and connections respectively. It isn't working in today's Tor, due to code rot and possible lingering libevent bugs. More work is needed; it would be good to get this working efficiently again, to have IOCP support on Windows.

Controlling connections

A connection can have reading or writing enabled or disabled for a wide variety of reasons, including:

  • Writing is disabled when there is no more data to write
  • For some connection types, reading is disabled when the inbuf is too full.
  • Reading/writing is temporarily disabled on connections that have recently read/written enough data up to their bandwidth
  • Reading is disabled on connections when reading more data from them would require that data to be buffered somewhere else that is already full.

Currently, these conditions are checked in a diffuse set of increasingly complex conditional expressions. In the future, it could be helpful to transition to a unified model for handling temporary read/write suspensions.

Kinds of connections

Today Tor has the following connection and pseudoconnection types. For the most part, each type of channel has an associated C module that implements its underlying logic.

Edge connections receive data from and deliver data to points outside the onion routing network. See connection_edge.c. They fall into two types:

Entry connections are a type of edge connection. They receive data from the user running a Tor client, and deliver data to that user. They are used to implement SOCKSPort, TransPort, NATDPort, and so on. Sometimes they are called "AP" connections for historical reasons (it used to stand for "Application Proxy").

Exit connections are a type of edge connection. They exist at an exit node, and transmit traffic to and from the network.

(Entry connections and exit connections are also used as placeholders when performing a remote DNS request; they are not decoupled from the notion of "stream" in the Tor protocol. This is implemented partially in connection_edge.c, and partially in dnsserv.c and dns.c.)

OR connections send and receive Tor cells over TLS, using some version of the Tor link protocol. Their implementation is spread across connection_or.c, with a bit of logic in command.c, relay.c, and channeltls.c.

Extended OR connections are a type of OR connection for use on bridges using pluggable transports, so that the PT can tell the bridge some information about the incoming connection before passing on its data. They are implemented in ext_orport.c.

Directory connections are server-side or client-side connections that implement Tor's HTTP-based directory protocol. These are instantiated using a socket when Tor is making an unencrypted HTTP connection. When Tor is tunneling a directory request over a Tor circuit, directory connections are implemented using a linked connection pair (see below). Directory connections are implemented in directory.c; some of the server-side logic is implemented in dirserver.c.

Controller connections are local connections to a controller process implementing the controller protocol from control-spec.txt. These are in control.c.

Listener connections are not stream oriented! Rather, they wrap a listening socket in order to detect new incoming connections. They bypass most of stream logic. They don't have associated buffers. They are implemented in connection.c.

structure hierarchy for connection types

>Note: "History Time!" You might occasionally find reference to a couple types of connections

which no longer exist in modern Tor. A CPUWorker connection >connected the main Tor process to a thread or process used for >computation. (Nowadays we use in-process communication.) Even more >anciently, a DNSWorker connection connected the main tor process to >a separate thread or process used for running gethostbyname() or >getaddrinfo(). (Nowadays we use Libevent's evdns facility to >perform DNS requests asynchronously.)

Linked connections

Sometimes two channels are joined together, such that data which the Tor process sends on one should immediately be received by the same Tor process on the other. (For example, when Tor makes a tunneled directory connection, this is implemented on the client side as a directory connection whose output goes, not to the network, but to a local entry connection. And when a directory receives a tunnelled directory connection, this is implemented as an exit connection whose output goes, not to the network, but to a local directory connection.)

The earliest versions of Tor to support linked connections used socketpairs for the purpose. But using socketpairs forced us to copy data through kernelspace, and wasted limited file descriptors. So instead, a pair of connections can be linked in-process. Each linked connection has a pointer to the other, such that data written on one is immediately readable on the other, and vice versa.

From connections to channels

There's an abstraction layer above OR connections (the ones that handle cells) and below cells called Channels. A channel's purpose is to transmit authenticated cells from one Tor instance (relay or client) to another.

Currently, only one implementation exists: Channel_tls, which sends and receiveds cells over a TLS-based OR connection.

Cells are sent on a channel using channel_write_{,packed_,var_}cell(). Incoming cells arrive on a channel from its backend using channel_queue*_cell(), and are immediately processed using channel_process_cells().

Some cell types are handled below the channel layer, such as those that affect handshaking only. And some others are passed up to the generic cross-channel code in command.c: cells like DESTROY and CREATED are all trivial to handle. But relay cells require special handling...

From channels through circuits

When a relay cell arrives on an existing circuit, it is handled in circuit_receive_relay_cell() – one of the innermost functions in Tor. This function encrypts or decrypts the relay cell as appropriate, and decides whether the cell is intended for the current hop of the circuit.

If the cell is intended for the current hop, we pass it to connection_edge_process_relay_cell() in relay.c, which acts on it based on its relay command, and (possibly) queues its data on an edge_connection_t.

If the cell is not intended for the current hop, we queue it for the next channel in sequence with append cell_to_circuit_queue(). This places the cell on a per-circuit queue for cells headed out on that particular channel.

Sending cells on circuits: the complicated bit.

Relay cells are queued onto circuits from one of two (main) sources: reading data from edge connections, and receiving a cell to be relayed on a circuit. Both of these sources place their cells on cell queue: each circuit has one cell queue for each direction that it travels.

A naive implementation would skip using cell queues, and instead write each outgoing relay cell. (Tor did this in its earlier versions.) But such an approach tends to give poor performance, because it allows high-volume circuits to clog channels, and it forces the Tor server to send data queued on a circuit even after that circuit has been closed.

So by using queues on each circuit, we can add cells to each channel on a just-in-time basis, choosing the cell at each moment based on a performance-aware algorithm.

This logic is implemented in two main modules: scheduler.c and circuitmux*.c. The scheduler code is responsible for determining globally, across all channels that could write cells, which one should next receive queued cells. The circuitmux code determines, for all of the circuits with queued cells for a channel, which one should queue the next cell.

(This logic applies to outgoing relay cells only; incoming relay cells are processed as they arrive.)