                    THE ANDROID "QEMUD" MULTIPLEXING DAEMON

I. Overview:
------------

The Android system image includes a small daemon program named "qemud"
which is started at boot time. Its purpose is to provide a multiplexing
communication channel between the emulated system and the emulator program
itself. Another way to support communication between the emulated system and
the emulator program is using qemu pipes (see ANDROID-QEMU-PIPE.TXT for details
on qemu pipes).

Its goal is to allow certain parts of the system to talk directly to the
emulator without requiring special kernel support; this simplifies a lot of
things since it does *not* require:

- writing/configuring a specific kernel driver
- writing the corresponding hardware emulation code in hw/goldfish_xxxx.c
- dealing with device allocation and permission issues in the emulated system

The emulator provides 'services' to various parts of the emulated system.
Each service is identified by a name and serves a specific purpose. For
example:

   "gsm"       Used to communicate with the emulated GSM modem with
               AT commands.

   "gps"       Used to receive NMEA sentences broadcasted from the
               emulated GPS device.

   "sensors"   Used to list the number of emulated sensors, as well as
               enable/disable reception of specific sensor events.

   "control"   Used to control misc. simple emulated hardware devices
               (e.g. vibrator, leds, LCD backlight, etc...)


II. Implementation:
-------------------

Since the "cupcake" platform, this works as follows:

- A 'qemud client' is any part of the emulated system that wants to talk
  to the emulator. It does so by:

  - connecting to the /dev/socket/qemud Unix domain socket
  - sending the service name through the socket
  - receives two bytes of data, which will be "OK" in case of
    success, or "KO" in case of failure.

  After an OK, the same connection can be used to talk directly to the
  corresponding service.


- The /dev/socket/qemud Unix socket is created by init and owned by the
  'qemud' daemon started at boot by /system/etc/init.goldfish.rc

  The daemon also opens an emulated serial port (e.g. /dev/ttyS1) and
  will pass all messages between clients and emulator services. Thus,
  everything looks like the following:


     emulator <==serial==> qemud <---> /dev/socket/qemud <-+--> client1
                                                           |
                                                           +--> client2

  A very simple multiplexing protocol is used on the serial connection:

           offset    size    description

               0       2     2-char hex string giving the destination or
                             source channel

               2       4     4-char hex string giving the payload size

               6       n     the message payload

  Where each client gets a 'channel' number allocated by the daemon
  at connection time.

  Note that packets larger than 65535 bytes cannot be sent directly
  through the qemud channel. This is intentional; for large data
  communication, the client and service should use a fragmentation
  convention that deals with this.

  Zero-sized packets are silently discard by qemud and the emulator and
  should normally not appear on the serial port.

  Channel 0 is reserved for control messages between the daemon and the
  emulator. These are the following:

    - When a client connects to /dev/socket/qemud and sends a service
      name to the daemon, the later sends to the emulator:

        connect:<service>:<id>

      where <service> is the service name, and <id> is a 2-hexchar string
      giving the allocated channel index for the client.


    - The emulator can respond in case of success with:

        ok:connect:<id>

      or, in case of failure, with:

        ok:connect:<id>:<reason>

      where <reason> is a liberal string giving the reason for failure.
      It is never sent to clients (which will only receive a "KO") and
      is used strictly for debugging purposes.

    - After a succesful connect, all messages between the client and
      the corresponding emulator service will be passed through the
      corresponding numbered channel.

      But if the client disconnects from the socket, the daemon will
      send through channel 0 this message to the emulator:

        disconnect:<id>

    - If an emulator service decides, for some reason, to disconnect
      a client, the emulator will send to the daemon (on channel 0):

        disconnect:<id>

      The daemon deals with this gracefully (e.g. it will wait that the
      client has read all buffered data in the daemon before closing the
      socket, to avoid packet loss).

    - Any other command sent from the daemon to the emulator will result
      in the following answer:

        ko:bad command

- Which exact serial port to open is determined by the emulator at startup
  and is passed to the system as a kernel parameter, e.g.:

        android.qemud=ttyS1


- The code to support services and their clients in the emulator is located
  in android/hw-qemud.c. This code is heavily commented.

  The daemon's source is in $ROOT/development/emulator/qemud/qemud.c

  The header in $ROOT/hardware/libhardware/include/hardware/qemud.h
  can be used by clients to ease connecting and talking to QEMUD-based
  services.

  This is used by $ROOT/developement/emulator/sensors/sensors_qemu.c which
  implements emulator-specific sensor support in the system by talking to
  the "sensors" service provided by the emulator (if available).

  Code in $ROOT/hardware/libhardware_legacy also uses QEMUD-based services.


- Certain services also implement a simple framing protocol when exchanging
  messages with their clients. The framing happens *after* serial port
  multiplexing and looks like:

           offset    size    description

               0       4     4-char hex string giving the payload size

               4       n     the message payload

  This is needed because the framing protocol used on the serial port is
  not preserved when talking to clients through /dev/socket/qemud.

  Certain services do not need it at all (GSM, GPS) so it is optional and
  must be used depending on which service you talk to by clients.

- QEMU pipe communication model works similarly to the serial port multiplexing,
  but also has some differences as far as connecting client with the service is
  concerned:

     emulator <-+--> /dev/qemu_pipe/qemud:srv1 <---> client1
                |
                +--> /dev/qemu_pipe/qemud:srv2 <---> client2

  In the pipe model each client gets connected to the emulator through a unique
  handle to /dev/qemu_pipe (a "pipe"), so there is no need for multiplexing the
  channels.

III. Legacy 'qemud':
--------------------

The system images provided by the 1.0 and 1.1 releases of the Android SDK
implement an older variant of the qemud daemon that uses a slightly
different protocol to communicate with the emulator.

This is documented here since this explains some subtleties in the
implementation code of android/hw-qemud.c

The old scheme also used a serial port to allow the daemon and the emulator
to communicate. However, the multiplexing protocol swaps the position of
'channel' and 'length' in the header:

           offset    size    description

               0       4     4-char hex string giving the payload size

               4       2     2-char hex string giving the destination or
                             source channel

               6       n     the message payload

Several other differences, best illustrated by the following graphics:

    emulator <==serial==> qemud <-+--> /dev/socket/qemud_gsm <--> GSM client
                                  |
                                  +--> /dev/socket/qemud_gps <--> GPS client
                                  |
                                  +--> /dev/socket/qemud_control <--> client(s)

Now, for the details:

 - instead of a single /dev/socket/qemud, init created several Unix domain
   sockets, one per service:

        /dev/socket/qemud_gsm
        /dev/socket/qemud_gps
        /dev/socket/qemud_control

   note that there is no "sensors" socket in 1.0 and 1.1

 - the daemon created a de-facto numbered channel for each one of these
   services, even if no client did connect to it (only one client could
   connect to a given service at a time).

 - at startup, the emulator does query the channel numbers of all services
   it implements, e.g. it would send *to* the daemon on channel 0:

        connect:<service>

   where <service> can be one of "gsm", "gps" or "control"

   (Note that on the current implementation, the daemon is sending connection
   messages to the emulator instead).

 - the daemon would respond with either:

        ok:connect:<service>:<hxid>

   where <service> would be the service name, and <hxid> a 4-hexchar channel
   number (NOTE: 4 chars, not 2). Or with:

        ko:connect:bad name


This old scheme was simpler to implement in both the daemon and the emulator
but lacked a lot of flexibility:

  - adding a new service required to modify /system/etc/init.goldfish.rc
    as well as the daemon source file (which contained a hard-coded list
    of sockets to listen to for client connections).

  - only one client could be connected to a given service at a time,
    except for the GPS special case which was a unidirectionnal broadcast
    by convention.

The current implementation moves any service-specific code to the emulator,
only uses a single socket and allows concurrent clients for a all services.


IV. State snapshots:
--------------------

Support for snapshots relies on the symmetric qemud_*_save and qemud_*_load
functions which save the state of the various Qemud* structs defined in
android/hw-qemud.c. The high-level process is as follows.

When a snapshot is made, the names and configurations of all services are
saved. Services can register a custom callback, which is invoked at this point
to allow saving of service-specific state. Next, clients are saved following
the same pattern. We save the channel id and the name of service they are
registered to, then invoke a client-specific callback.

When a snapshot is restored, the first step is to check whether all services
that were present when the snapshot was made are available. There is currently
no functionality to start start missing services, so loading fails if a service
is not present. If all services are present, callbacks are used to restore
service-specific state.

Next, all active clients are shut down. Information from the snapshot is used
to start new clients for the services and channels as they were when the
snapshot was made. This completes the restore process.