Tor  0.4.7.0-alpha-dev
Initialization and shutdown

Overview

Tor has a single entry point: tor_run_main() in main.c. All the ways of starting a Tor process (ntmain.c, tor_main.c, and tor_api.c) work by invoking tor_run_main().

The tor_run_main() function normally exits (1) by returning: not by calling abort() or exit(). Before it returns, it calls tor_cleanup() in shutdown.c.

Conceptually, there are several stages in running Tor.

  1. First, we initialize those modules that do not depend on the configuration. This happens in the first half of tor_run_main(), and the first half of tor_init(). (2)
  2. Second, we parse the command line and our configuration, and configure systems that depend on our configuration or state. This configuration happens midway through tor_init(), which invokes options_init_from_torrc(). We then initialize more systems from the second half of tor_init().
  3. At this point we may exit early if we have been asked to do something requiring no further initialization, like printing our version number or creating a new signing key. Otherwise, we proceed to run_tor_main_loop(), which initializes some network-specific parts of Tor, grabs some daemon-only resources (like the data directory lock) and starts Tor itself running.

1. tor_run_main() can terminate with a call to abort() or exit(), but only when crashing due to a bug, or when forking to run as a daemon.

2. The pieces of code that I'm describing as "the first part of tor_init()" and so on deserve to be functions with their own name. I'd like to refactor them, but before I do so, there is some slight reorganization that needs to happen. Notably, the nt_service_parse_options() call ought logically to be later in our initialization sequence. See ticket 32447 for our refactoring progress.

Subsystems and initialization

Our current convention is to use the subsystem mechanism to initialize and clean up pieces of Tor. The more recently updated pieces of Tor will use this mechanism. For examples, see e.g. time_sys.c or log_sys.c.

In simplest terms, a subsystem is a logically separate part of Tor that can be initialized, shut down, managed, and configured somewhat independently of the rest of the program.

The subsys_fns_t type describes a subsystem and a set of functions that initialize it, desconstruct it, and so on. To define a subsystem, we declare a const instance of subsys_fns_t. See the documentation for subsys_fns_t for a full list of these functions.

After defining a subsystem, it must be inserted in subsystem_list.c. At that point, table-driven mechanisms in subsysmgr.c will invoke its functions when appropriate.

Initialization versus configuration

We note that the initialization phase of Tor occurs before any configuration is read from disk – and therefore before any other files are read from disk. Therefore, any behavior that depends on Tor's configuration or state must occur after the initialization process, during configuration.