Author: Arvid Norberg,

libtorrent plugins

libtorrent has a plugin interface for implementing extensions to the protocol. These can be general extensions for transferring metadata or peer exchange extensions, or it could be used to provide a way to customize the protocol to fit a particular (closed) network.

In short, the plugin interface makes it possible to:

  • register extension messages (sent in the extension handshake), see extensions.
  • add data and parse data from the extension handshake.
  • send extension messages and standard bittorrent messages.
  • override or block the handling of standard bittorrent messages.
  • save and restore state via the session state
  • see all alerts that are posted

a word of caution

Writing your own plugin is a very easy way to introduce serious bugs such as dead locks and race conditions. Since a plugin has access to internal structures it is also quite easy to sabotage libtorrent's operation.

All the callbacks in this interface are called with the main libtorrent thread mutex locked. And they are always called from the libtorrent network thread. In case portions of your plugin are called from other threads, typically the main thread, you cannot use any of the member functions on the internal structures in libtorrent, since those require the mutex to be locked. Futhermore, you would also need to have a mutex on your own shared data within the plugin, to make sure it is not accessed at the same time from the libtorrent thread (through a callback). See boost thread's mutex. If you need to send out a message from another thread, it is advised to use an internal queue, and do the actual sending in tick().

Since the plugin interface gives you easy access to internal structures, it is not supported as a stable API. Plugins should be considered spcific to a specific version of libtorrent. Although, in practice the internals mostly don't change that dramatically.

plugin interface

The plugin interface consists of three base classes that the plugin may implement. These are called plugin, torrent_plugin and peer_plugin. They are found in the <libtorrent/extensions.hpp> header.

These plugins are instantiated for each session, torrent and possibly each peer, respectively.

For plugins that only need per torrent state, it is enough to only implement torrent_plugin and pass a constructor function or function object to session::add_extension() or torrent_handle::add_extension() (if the torrent has already been started and you want to hook in the extension at run-time).

The signature of the function is:

boost::shared_ptr<torrent_plugin> (*)(torrent*, void*);

The first argument is the internal torrent object, the second argument is the userdata passed to session::add_torrent() or torrent_handle::add_extension().

The function should return a boost::shared_ptr<torrent_plugin> which may or may not be 0. If it is a null pointer, the extension is simply ignored for this torrent. If it is a valid pointer (to a class inheriting torrent_plugin), it will be associated with this torrent and callbacks will be made on torrent events.

For more elaborate plugins which require session wide state, you would implement plugin, construct an object (in a boost::shared_ptr) and pass it in to session::add_extension().


struct plugin
        virtual ~plugin();
        virtual boost::shared_ptr<torrent_plugin> new_torrent(torrent* t, void* user);

        virtual void added(boost::weak_ptr<aux::session_impl> s);
        virtual void on_alert(alert const* a);
        virtual void on_tick();
        virtual void save_state(entry& ent) const;
        virtual void load_state(lazy_entry const& ent);


The synopsis for torrent_plugin follows:

struct torrent_plugin
        virtual ~torrent_plugin();
        virtual boost::shared_ptr<peer_plugin> new_connection(peer_connection*);

        virtual void on_piece_pass(int index);
        virtual void on_piece_failed(int index);

        virtual void tick();

        virtual bool on_pause();
        virtual bool on_resume();

        virtual void on_files_checked();

        virtual void on_state(int s);

        enum flags_t {
                first_time = 1,
                filtered = 2

        virtual void on_add_peer(tcp::endpoint const& ip
                , int src, int flags);

This is the base class for a torrent_plugin. Your derived class is (if added as an extension) instantiated for each torrent in the session. The callback hook functions are defined as follows.


boost::shared_ptr<peer_plugin> new_connection(peer_connection*);

This function is called each time a new peer is connected to the torrent. You may choose to ignore this by just returning a default constructed shared_ptr (in which case you don't need to override this member function).

If you need an extension to the peer connection (which most plugins do) you are supposed to return an instance of your peer_plugin class. Which in turn will have its hook functions called on event specific to that peer.

The peer_connection will be valid as long as the shared_ptr is being held by the torrent object. So, it is generally a good idea to not keep a shared_ptr to your own peer_plugin. If you want to keep references to it, use weak_ptr.

If this function throws an exception, the connection will be closed.

on_piece_pass() on_piece_fail()

void on_piece_pass(int index);
void on_piece_failed(int index);

These hooks are called when a piece passes the hash check or fails the hash check, respectively. The index is the piece index that was downloaded. It is possible to access the list of peers that participated in sending the piece through the torrent and the piece_picker.


void tick();

This hook is called approximately once per second. It is a way of making it easy for plugins to do timed events, for sending messages or whatever.

on_pause() on_resume()

bool on_pause();
bool on_resume();

These hooks are called when the torrent is paused and unpaused respectively. The return value indicates if the event was handled. A return value of true indicates that it was handled, and no other plugin after this one will have this hook function called, and the standard handler will also not be invoked. So, returning true effectively overrides the standard behavior of pause or unpause.

Note that if you call pause() or resume() on the torrent from your handler it will recurse back into your handler, so in order to invoke the standard handler, you have to keep your own state on whether you want standard behavior or overridden behavior.


void on_files_checked();

This function is called when the initial files of the torrent have been checked. If there are no files to check, this function is called immediately.

i.e. This function is always called when the torrent is in a state where it can start downloading.


enum flags_t {
        first_time = 1,
        filtered = 2

virtual void on_add_peer(tcp::endpoint const& ip
        , int src, int flags);

This function is called whenever we hear about a peer from any peer source, such as the tracker, PEX, DHT or Local peer discovery.

src is a bitmask of peer_info::peer_source_flags:

enum peer_source_flags
        tracker = 0x1,
        dht = 0x2,
        pex = 0x4,
        lsd = 0x8,
        resume_data = 0x10,
        incoming = 0x20

flags is a bitmask of:

enum flags_t {
        first_time = 1,
        filtered = 2

If the filtered flag is set, it means the peer wasn't added to the peer list because of and IP filter, port filter, reserved ports filter.


struct peer_plugin
        virtual ~peer_plugin();

        virtual void add_handshake(entry&);
        virtual bool on_handshake(char const* reserved_bits);
        virtual bool on_extension_handshake(lazy_entry const& h);

        virtual bool on_choke();
        virtual bool on_unchoke();
        virtual bool on_interested();
        virtual bool on_not_interested();
        virtual bool on_have(int index);
        virtual bool on_bitfield(bitfield const& bits);
        virtual bool on_have_all();
        virtual bool on_have_none();
        virtual bool on_allowed_fast(int index);
        virtual bool on_request(peer_request const& req);
        virtual bool on_piece(peer_request const& piece, disk_buffer_holder& buffer);
        virtual bool on_cancel(peer_request const& req);
        virtual bool on_reject(peer_request const& req);
        virtual bool on_suggest(int index);
        virtual bool on_extended(int length
                , int msg, buffer::const_interval body);
        virtual bool on_unknown_message(int length, int msg
                , buffer::const_interval body);
        virtual void on_piece_pass(int index);
        virtual void on_piece_failed(int index);

        virtual void tick();

        virtual bool write_request(peer_request const& r);


struct disk_buffer_holder
        disk_buffer_holder(aux::session_impl& s, char* b);
        char* release();
        char* buffer();

The disk buffer holder acts like a scoped_ptr that frees a disk buffer when it's destructed, unless it's released. release returns the disk buffer and transferres ownership and responsibility to free it to the caller.

A disk buffer is freed by passing it to session_impl::free_disk_buffer().

buffer() returns the pointer without transferring responsibility. If this buffer has been released, buffer() will return 0.

custom alerts

Since plugins are running within internal libtorrent threads, one convenient way to communicate with the client is to post custom alerts.

The expected interface of any alert, apart from deriving from the alert base class, looks like this:

const static int alert_type = <unique alert ID>;
virtual int type() const { return alert_type; }

virtual std::string message() const;

virtual std::auto_ptr<alert> clone() const
{ return std::auto_ptr<alert>(new name(*this)); }

const static int static_category = <bitmask of alert::category_t flags>;
virtual int category() const { return static_category; }

virtual char const* what() const { return <string literal of the name of this alert>; }

The alert_type is used for the type-checking in alert_cast. It must not collide with any other alert. The built-in alerts in libtorrent will not use alert type IDs greater than user_alert_id. When defining your own alert, make sure it's greater than this constant.

type() is the run-time equivalence of the alert_type.

The message() virtual function is expected to construct a useful string representation of the alert and the event or data it represents. Something convenient to put in a log file for instance.

clone() is used internally to copy alerts. The suggested implementation of simply allocating a new instance as a copy of *this is all that's expected.

The static category is required for checking wether or not the category for a specific alert is enabled or not, without instantiating the alert. The category virtual function is the run-time equivalence.

The what() virtual function may simply be a string literal of the class name of your alert.

For more information, see the alert section in the main manual.