protocol.txt

--- a replacement for aproto -------------------------------------------

When it comes down to it, aproto's primary purpose is to forward
various streams between the host computer and client device (in either
direction).

This replacement further simplifies the concept, reducing the protocol
to an extremely straightforward model optimized to accomplish the
forwarding of these streams and removing additional state or
complexity.

The host side becomes a simple comms bridge with no "UI", which will 
be used by either commandline or interactive tools to communicate with 
a device or emulator that is connected to the bridge.

The protocol is designed to be straightforward and well-defined enough 
that if it needs to be reimplemented in another environment (Java 
perhaps), there should not problems ensuring perfect interoperability.

The protocol discards the layering aproto has and should allow the 
implementation to be much more robust.


--- protocol overview and basics ---------------------------------------

The transport layer deals in "messages", which consist of a 24 byte
header followed (optionally) by a payload.  The header consists of 6
32 bit words which are sent across the wire in little endian format.

struct message {
    unsigned command;       /* command identifier constant (A_CNXN, ...) */
    unsigned arg0;          /* first argument                            */
    unsigned arg1;          /* second argument                           */
    unsigned data_length;   /* length of payload (0 is allowed)          */
    unsigned data_crc32;    /* crc32 of data payload                     */
    unsigned magic;         /* command ^ 0xffffffff                      */
};

Receipt of an invalid message header, corrupt message payload, or an
unrecognized command MUST result in the closing of the remote
connection.  The protocol depends on shared state and any break in the
message stream will result in state getting out of sync.

The following sections describe the six defined message types in
detail.  Their format is COMMAND(arg0, arg1, payload) where the payload
is represented by a quoted string or an empty string if none should be
sent.

The identifiers "local-id" and "remote-id" are always relative to the
*sender* of the message, so for a receiver, the meanings are effectively
reversed.



--- CONNECT(version, maxdata, "system-identity-string") ----------------

Command constant: A_CNXN

The CONNECT message establishes the presence of a remote system.
The version is used to ensure protocol compatibility and maxdata
declares the maximum message body size that the remote system
is willing to accept.

Currently, version=0x01000000 and maxdata=256*1024. Older versions of adb
hard-coded maxdata=4096, so CONNECT and AUTH packets sent to a device must not
be larger than that because they're sent before the CONNECT from the device
that tells the adb server what maxdata the device can support.

Both sides send a CONNECT message when the connection between them is
established.  Until a CONNECT message is received no other messages may
be sent. Any messages received before a CONNECT message MUST be ignored.

If a CONNECT message is received with an unknown version or insufficiently
large maxdata value, the connection with the other side must be closed.

The system identity string should be "<systemtype>:<serialno>:<banner>"
where systemtype is "bootloader", "device", or "host", serialno is some
kind of unique ID (or empty), and banner is a human-readable version
or identifier string.  The banner is used to transmit useful properties.

--- STLS(type, version, "") --------------------------------------------

Command constant: A_STLS

The TLS message informs the recipient that the connection will be encrypted
and will need to perform a TLS handshake. version is the current version of
the protocol.


--- AUTH(type, 0, "data") ----------------------------------------------

Command constant: A_AUTH

The AUTH message informs the recipient that authentication is required to
connect to the sender. If type is TOKEN(1), data is a random token that
the recipient can sign with a private key. The recipient replies with an
AUTH packet where type is SIGNATURE(2) and data is the signature. If the
signature verification succeeds, the sender replies with a CONNECT packet.

If the signature verification fails, the sender replies with a new AUTH
packet and a new random token, so that the recipient can retry signing
with a different private key.

Once the recipient has tried all its private keys, it can reply with an
AUTH packet where type is RSAPUBLICKEY(3) and data is the public key. If
possible, an on-screen confirmation may be displayed for the user to
confirm they want to install the public key on the device.


--- OPEN(local-id, 0, "destination") -----------------------------------

Command constant: A_OPEN

The OPEN message informs the recipient that the sender has a stream
identified by local-id that it wishes to connect to the named
destination in the message payload.  The local-id may not be zero.

The OPEN message MUST result in either a READY message indicating that
the connection has been established (and identifying the other end) or
a CLOSE message, indicating failure.  An OPEN message also implies
a READY message sent at the same time.

Common destination naming conventions include:

* "tcp:<host>:<port>" - host may be omitted to indicate localhost
* "udp:<host>:<port>" - host may be omitted to indicate localhost
* "vsock:<CID>:<port>"
* "local-dgram:<identifier>"
* "local-stream:<identifier>"
* "shell" - local shell service
* "upload" - service for pushing files across (like aproto's /sync)
* "fs-bridge" - FUSE protocol filesystem bridge


--- READY(local-id, remote-id, "") -------------------------------------

Command constant: A_OKAY

The READY message informs the recipient that the sender's stream
identified by local-id is ready for write messages and that it is
connected to the recipient's stream identified by remote-id.

Neither the local-id nor the remote-id may be zero.

A READY message containing a remote-id which does not map to an open
stream on the recipient's side is ignored.  The stream may have been
closed while this message was in-flight.

The local-id is ignored on all but the first READY message (where it
is used to establish the connection).  Nonetheless, the local-id MUST
not change on later READY messages sent to the same stream.


--- WRITE(local-id, remote-id, "data") ---------------------------------

Command constant: A_WRTE

The WRITE message sends data to the recipient's stream identified by
remote-id.  The payload MUST be <= maxdata in length.

A WRITE message containing a remote-id which does not map to an open
stream on the recipient's side is ignored.  The stream may have been
closed while this message was in-flight.

A WRITE message may not be sent until a READY message is received.
Once a WRITE message is sent, an additional WRITE message may not be
sent until another READY message has been received.  Recipients of
a WRITE message that is in violation of this requirement will CLOSE
the connection.


--- CLOSE(local-id, remote-id, "") -------------------------------------

Command constant: A_CLSE

The CLOSE message informs recipient that the connection between the
sender's stream (local-id) and the recipient's stream (remote-id) is
broken.  The remote-id MUST not be zero, but the local-id MAY be zero
if this CLOSE indicates a failed OPEN.

A CLOSE message containing a remote-id which does not map to an open
stream on the recipient's side is ignored.  The stream may have
already been closed by the recipient while this message was in-flight.

The recipient should not respond to a CLOSE message in any way.  The
recipient should cancel pending WRITEs or CLOSEs, but this is not a
requirement, since they will be ignored.


--- SYNC(online, sequence, "") -----------------------------------------

Command constant: A_SYNC

*** obsolete, no longer used ***

The SYNC message was used by the io pump to make sure that stale
outbound messages are discarded when the connection to the remote side
is broken.  It was only used internally to the bridge and never valid
to send across the wire.

* when the connection to the remote side goes offline, the io pump
  sends a SYNC(0, 0) and starts discarding all messages
* when the connection to the remote side is established, the io pump
  sends a SYNC(1, token) and continues to discard messages
* when the io pump receives a matching SYNC(1, token), it once again
  starts accepting messages to forward to the remote side


--- message command constants ------------------------------------------

#define A_SYNC 0x434e5953
#define A_CNXN 0x4e584e43
#define A_AUTH 0x48545541
#define A_OPEN 0x4e45504f
#define A_OKAY 0x59414b4f
#define A_CLSE 0x45534c43
#define A_WRTE 0x45545257
#define A_STLS 0x534C5453



--- implementation details ---------------------------------------------

The core of the bridge program will use three threads.  One thread
will be a select/epoll loop to handle io between various inbound and
outbound connections and the connection to the remote side.

The remote side connection will be implemented as two threads (one for
reading, one for writing) and a datagram socketpair to provide the
channel between the main select/epoll thread and the remote connection
threadpair.  The reason for this is that for usb connections, the
kernel interface on linux and osx does not allow you to do meaningful
nonblocking IO.

The endian swapping for the message headers will happen (as needed) in
the remote connection threadpair and that the rest of the program will
always treat message header values as native-endian.

The bridge program will be able to have a number of mini-servers
compiled in.  They will be published under known names (examples
"shell", "fs-bridge", etc) and upon receiving an OPEN() to such a
service, the bridge program will create a stream socketpair and spawn
a thread or subprocess to handle the io.


--- simplified / embedded implementation -------------------------------

For limited environments, like the bootloader, it is allowable to
support a smaller, fixed number of channels using pre-assigned channel
ID numbers such that only one stream may be connected to a bootloader
endpoint at any given time.  The protocol remains unchanged, but the
"embedded" version of it is less dynamic.

The bootloader will support two streams.  A "bootloader:debug" stream,
which may be opened to get debug messages from the bootloader and a 
"bootloader:control", stream which will support the set of basic 
bootloader commands.

Example command stream dialogues:  
  "flash_kernel,2515049,........\n" "okay\n" 
  "flash_ramdisk,5038,........\n" "fail,flash write error\n" 
  "bogus_command......" <CLOSE>


--- future expansion ---------------------------------------------------

I plan on providing either a message or a special control stream so that
the client device could ask the host computer to setup inbound socket
translations on the fly on behalf of the client device.


The initial design does handshaking to provide flow control, with a
message flow that looks like:

  >OPEN <READY >WRITE <READY >WRITE <READY >WRITE <CLOSE

The far side may choose to issue the READY message as soon as it receives
a WRITE or it may defer the READY until the write to the local stream
succeeds.  A future version may want to do some level of windowing where
multiple WRITEs may be sent without requiring individual READY acks.

------------------------------------------------------------------------

--- smartsockets -------------------------------------------------------

Port 5037 is used for smart sockets which allow a client on the host
side to request access to a service in the host adb daemon or in the
remote (device) daemon.  The service is requested by ascii name,
preceeded by a 4 digit hex length.  Upon successful connection an
"OKAY" response is sent, otherwise a "FAIL" message is returned.  Once
connected the client is talking to that (remote or local) service.

client: <hex4> <service-name>
server: "OKAY"

client: <hex4> <service-name>
server: "FAIL" <hex4> <reason>