transmission/third-party/dht

The files dht.c and dht.h implement the variant of the Kademlia Distributed
Hash Table (DHT) used in the Bittorrent network (``mainline'' variant).

The file dht-example.c is a stand-alone program that participates in the
DHT.  Another example is a patch against Transmission, which you might or
might not be able to find somewhere.

The code is designed to work well in both event-driven and threaded code.
The caller, which is either an event-loop or a dedicated thread, must
periodically call the function dht_periodic.  In addition, it must call
dht_periodic whenever any data has arrived from the network.

All functions return -1 in case of failure, in which case errno is set, or
a positive value in case of success.

Initialisation
**************

* dht_init

This must be called before using the library.  You pass it a bound IPv4
datagram socket, and your node id, a 20-octet array that should be globally
unique.

Node ids must be well distributed, so you cannot just use your Bittorrent
id; you should either generate a truly random value (using plenty of
entropy), or at least take the SHA-1 of something.  However, it is a good
idea to keep the id stable, so you may want to store it in stable storage
at client shutdown.
 
* dht_uninit

This may be called at the end of the session.  If dofree is true, it frees
all the memory allocated for the DHT.  If dofree is false, this function
currently does nothing.

Bootstrapping
*************

The DHT needs to be taught a small number of contacts to begin functioning.
You can hard-wire a small number of stable nodes in your application, but
this obviously fails to scale.  You may save the list of known good nodes
at shutdown, and restore it at restart.  You may also grab nodes from
torrent files (the nodes field), and you may exchange contacts with other
Bittorrent peers using the PORT extension.

* dht_ping

This is the main bootstrapping primitive.  You pass it an address at which
you believe that a DHT node may be living, and a query will be sent.  If
a node replies, and if there is space in the routing table, it will be
inserted.

* dht_insert_node

This is a softer bootstrapping method, which doesn't actually send
a query -- it only stores the node in the routing table for later use.  It
is a good idea to use that when e.g. restoring your routing table from
disk.

Note that dht_insert_node requires that you supply a node id.  If the id
turns out to be wrong, the DHT will eventually recover; still, inserting
massive amounts of incorrect information into your routing table is
certainly not a good idea.

An additionaly difficulty with dht_insert_node is that, for various
reasons, a Kademlia routing table cannot absorb nodes faster than a certain
rate.  Dumping a large number of nodes into a table using dht_insert_node
will probably cause most of these nodes to be discarded straight away.
(The tolerable rate is difficult to estimate; it is probably on the order
of one node every few seconds per node already in the table divided by 8,
for some suitable value of 8.)

Doing some work
***************

* dht_periodic

This function should be called by your main loop periodically, and also
whenever data is available on the socket.  The time after which
dht_periodic should be called if no data is available is returned in the
parameter tosleep.  (You do not need to be particularly accurate; actually,
it is a good idea to be late by a random value.)

The parameter available indicates whether any data is available on the
socket.  If it is 0, dht_periodic will not try to read data; if it is 1, it
will.

Dht_periodic also takes a callback, which will be called whenever something
interesting happens (see below).

* dht_search

This schedules a search for information about the info-hash specified in
id.  If port is not 0, it specifies the TCP port on which the current peer
is litening; in that case, when the search is complete it will be announced
to the network.  The port is in host order, beware if you got it from
a struct sockaddr_in.

In either case, data is passed to the callback function as soon as it is
available, possibly in multiple pieces.  The callback function will
additionally be called when the search is complete.

Up to DHT_MAX_SEARCHES (20) searches can be in progress at a given time;
any more, and dht_search will return -1.  If you specify a new search for
the same info hash as a search still in progress, the previous search is
combined with the new one -- you will only receive a completion indication
once.

Information queries
*******************

* dht_nodes

This returns the number of known good, dubious and cached nodes in our
routing table.  This can be used to decide whether it's reasonable to start
a search; a search is likely to be successful as long as we have a few good
nodes; however, in order to avoid overloading your bootstrap nodes, you may
want to wait until good is at least 4 and good + doubtful is at least 30 or
so.

It also includes the number of nodes that recently send us an unsolicited
request; this can be used to determine if the UDP port used for the DHT is
firewalled.

If you want to display a single figure to the user, you should display good
+ doubtful, which is the total number of nodes in your routing table.  Some
clients try to estimate the total number of nodes, but this doesn't make
much sense -- since the result is exponential in the number of nodes in the
routing table, small variations in the latter cause huge jumps in the
former.

* dht_get_nodes

This retrieves the list of known good nodes, starting with the nodes in our
own bucket.  It is a good idea to save the list of known good nodes at
shutdown, and ping them at startup.

* dht_dump_tables
* dht_debug

These are debugging aids.

Functions provided by you
*************************

* The callback function

The callback function is called with 5 arguments.  Closure is simply the
value that you passed to dht_periodic.  Event is one of DHT_EVENT_VALUES,
which indicates that we have new values, or DHT_EVENT_SEARCH_DONE, which
indicates that a search has completed.  In either case, info_hash is set to
the info-hash of the search.

In the case of DHT_EVENT_VALUES, data is a list of nodes in ``compact''
format -- 6 bytes per node, 4 for the IP address and 2 for the port.  It's
length in bytes is in data_len.

* dht_hash

This should compute a reasonably strong cryptographic hash of the passed
values.  It should map cleanly to your favourite crypto toolkit's MD5 or
SHA-1 function.

Final notes
***********

* NAT

Nothing works well across NATs, but Kademlia is somewhat less impacted than
many other protocols.  The implementation takes care to distinguish between
unidirectional and bidirectional reachability, and NATed nodes will
eventually fall out from other nodes' routing tables.

While there is no periodic pinging in this implementation, maintaining
a full routing table requires slightly more than one packet exchange per
minute, even in a completely idle network; this should be sufficient to
make most full cone NATs happy.

* Missing functionality

Some of the code has had very little testing.  If it breaks, you get to
keep both pieces.

There is currently no good way to save and restore your routing table.

IPv6 support is deliberately not included: designing a double-stack
distributed hash table raises some tricky issues, and doing it naively may
break connectivity for everyone.

                                        Juliusz Chroboczek
                                        <jch@pps.jussieu.fr>