csprotocol.txt

This is kind of informal and may be wrong, but it helped me.  It's
basically a summary of clientserver.c and authenticate.c.

 -- Martin Pool <mbp@samba.org>


This is the protocol used for rsync --daemon; i.e. connections to port
873 rather than invocations over a remote shell.

When the server accepts a connection, it prints a newline-terminated
greeting line:

  @RSYNCD: <version>.<subprotocol> <digest1> <digestN>

The <version> is the numeric version (see PROTOCOL_VERSION in rsync.h)
The <subprotocol> is the numeric subprotocol version (which is 0 for a
final protocol version, as the SUBPROTOCOL_VERSION define discusses).
The <digestN> names are the authentication digest algorithms that the
daemon supports, listed in order of preference.

An rsync prior to 3.2.7 omits the digest names.  An rsync prior to 3.0.0
also omits the period and the <subprotocol> value.  Since a final
protocol has a subprotocol value of 0, a missing subprotocol value is
assumed to be 0 for any protocol prior to 30.  It is considered a fatal
error for protocol 30 and above to omit it.  It is considered a fatal
error for protocol 32 and above to omit the digest name list (currently
31 is the newest protocol).

The daemon expects to see a similar greeting line back from the client.
Once received, the daemon follows the opening line with a free-format
text message-of-the-day (if any is defined).

The server is now in the connected state.  The client can either send
the command:

  #list

(to get a listing of modules) or the name of a module.  After this, the
connection is now bound to a particular module.  Access per host for
this module is now checked, as is per-module connection limits.

If authentication is required to use this module, the server will say:

  @RSYNCD: AUTHREQD <challenge>

where <challenge> is a random string of base64 characters.  The client
must respond with:

  <user> <response>

The <user> is the username they claim to be. The <response> is the
base64 form of the digest hash of the challenge+password string. The
chosen digest method is the most preferred client method that is also in
the server's list.  If no digest list was explicitly provided, the side
expecting a list assumes the other side provided either the single name
"md5" (for a negotiated protocol 30 or 31), or the single name "md4"
(for an older protocol).

At this point the server applies all remaining constraints before
handing control to the client, including switching uid/gid, setting up
include and exclude lists, moving to the root of the module, and doing
chroot.

If the login is acceptable, then the server will respond with

  @RSYNCD: OK

The client now writes some rsync options, as if it were remotely
executing the command.  The server parses these arguments as if it had
just been invoked with them, but they're added to the existing state.
So if the client specifies a list of files to be included or excluded,
they'll defer to existing limits specified in the server
configuration.

At this point the client and server both switch to using a
multiplexing layer across the socket.  The main point of this is to
allow the server to asynchronously pass errors back, while still
allowing streamed and pipelined data.

Unfortunately, the multiplex protocol is not used at every stage.  We
start up in plain socket mode and then change over by calling
io_start_buffering.  Of course both the client and the server have to
do this at the same point.

The server then talks to the client as normal across the socket,
passing checksums, file lists and so on.  For documentation of that,
stay tuned (or write it yourself!).



------------
Protocol version changes

31	(2013-09-28, 3.1.0)

	Initial release of protocol 31 had no changes.  Rsync 3.2.7
	introduced the suffixed list of digest names on the greeting
	line.  The presence of the list is allowed even if the greeting
	indicates an older protocol version number.

30	(2007-10-04, 3.0.0pre1)

	The use of a ".<subprotocol>" number was added to
	@RSYNCD: <version>.<subprotocol>

25	(2001-08-20, 2.4.7pre2) 

	Send an explicit "@RSYNC EXIT" command at the end of the
	module listing.  We never intentionally end the transmission
	by just closing the socket anymore.