path: root/dcop/HOWTO

		DCOP: Desktop COmmunications Protocol

		    Preston Brown <[email protected]>
			   October 14, 1999

	 Revised and extended by Matthias Ettrich <[email protected]>
			   Mar 29, 2000

        Extended with DCOP Signals by Waldo Bastian <[email protected]>
                           Feb 19, 2001


Motivation and Background:
--------------------------

The motivation behind building a protocol like DCOP is simple.  For
the past year, we have been attempting to enable interprocess
communication between KDE applications. KDE already has an extremely
simple IPC mechanism called KWMcom, which is (was!) used for communicating
between the panel and the window manager for instance.  It is about as
simple as it gets, passing messages via X Atoms.  For this reason it
is limited in the size and complexity of the data that can be passed
(X atoms must be small to remain efficient) and it also makes it so
that X is required.  CORBA was thought to be a more effective IPC/RPC
solution.  However, after a year of attempting to make heavy use of
CORBA in KDE, we have realized that it is a bit slow and memory
intensive for simple use.  It also has no authentication available.

What we really needed was an extremely simple protocol with basic
authorization, along the lines of MIT-MAGIC-COOKIE, as used by X.  It
would not be able to do NEARLY what CORBA was able to do, but for the
simple tasks required it would be sufficient. Some examples of such
tasks might be an application sending a message to the panel saying,
"I have started, stop displaying the 'application starting' wait
state," or having a new application that starts query to see if any
other applications of the same name are running.  If they are, simply
call a function on the remote application to create a new window,
rather than starting a new process.

Implementation:
---------------

DCOP is a simple IPC/RPC mechanism built to operate over sockets.
Either unix domain sockets or tcp/ip sockets are supported. DCOP is
built on top of the Inter Client Exchange (ICE) protocol, which comes
standard as a part of X11R6 and later. It also depends on Qt, but
beyond that it does not require any other libraries. Because of this,
it is extremely lightweight, enabling it to be linked into all KDE
applications with low overhead.

Model:
------

The model is simple.  Each application using DCOP is a client.  They
communicate to each other through a DCOP server, which functions like
a traffic director, dispatching messages/calls to the proper
destinations.  All clients are peers of each other.

Two types of actions are possible with DCOP: "send and forget"
messages, which do not block, and "calls," which block waiting for
some data to be returned.

Any data that will be sent is serialized (marshalled, for you CORBA
types) using the built-in QDataStream operators available in all of
the Qt classes.  This is fast and easy.  In fact it's so little work
that you can easily write the marshalling code by hand. In addition,
there's a simple IDL-like compiler available (dcopidl and dcopidl2cpp)
that generates stubs and skeletons for you. Using the dcopidl compiler
has the additional benefit of type safety.

This HOWTO describes the manual method first and covers the dcopidl
compiler later.

Establishing the Connection:
----------------------------

KApplication has gained a method called "KApplication::dcopClient()"
which returns a pointer to a DCOPClient instance.  The first time this
method is called, the client class will be created.  DCOPClients have
unique identifiers attached to them which are based on what
KApplication::name() returns.  In fact, if there is only a single
instance of the program running, the appId will be equal to
KApplication::name().

To actually enable DCOP communication to begin, you must use
DCOPClient::attach().  This will attempt to attach to the DCOP server.
If no server is found or there is any other type of error, attach()
will return false. KApplication will catch a dcop signal and display an
appropriate error message box in that case.

After connecting with the server via DCOPClient::attach(), you need to
register this appId with the server so it knows about you.  Otherwise,
you are communicating anonymously.  Use the
DCOPClient::registerAs(const QCString &name) to do so.  In the simple
case:

/*
 * returns the appId that is actually registered, which _may_ be
 * different from what you passed
 */
appId = client->registerAs(kApp->name());

If you never retrieve the DCOPClient pointer from KApplication, the
object will not be created and thus there will be no memory overhead.

You may also detach from the server by calling DCOPClient::detach().
If you wish to attach again you will need to re-register as well.  If
you only wish to change the ID under which you are registered, simply
call DCOPClient::registerAs() with the new name.

KUniqueApplication automatically registers itself to DCOP. If you
are using KUniqueApplication you should not attach or register
yourself, this is already done. The appId is by definition
equal to kapp->name(). You can retrieve the registered DCOP client
by calling kapp->dcopClient().

Sending Data to a Remote Application:
-------------------------------------

To actually communicate, you have one of two choices.  You may either
call the "send" or the "call" method.  Both methods require three
identification parameters: an application identifier, a remote object,
a remote function. Sending is asynchronous (i.e. it returns immediately)
and may or may not result in your own application being sent a message at
some point in the future. Then "send" requires one and "call" requires
two data parameters.

The remote object must be specified as an object hierarchy.  That is,
if the toplevel object is called "fooObject" and has the child
"barObject", you would reference this object as "fooObject/barObject".
Functions must be described by a full function signature.  If the
remote function is called "doIt", and it takes an int, it would be
described as "doIt(int)".  Please note that the return type is not
specified here, as it is not part of the function signature (or at
least the C++ understanding of a function signature).  You will get
the return type of a function back as an extra parameter to
DCOPClient::call().  See the section on call() for more details.

In order to actually get the data to the remote client, it must be
"serialized" via a QDataStream operating on a QByteArray. This is how
the data parameter is "built". A few examples will make clear how this
works.

Say you want to call "doIt" as described above, and not block (or wait
for a response).  You will not receive the return value of the remotely
called function, but you will not hang while the RPC is processed either.
The return value of send() indicates whether DCOP communication succeeded
or not.

QByteArray data;
QDataStream arg(data, IO_WriteOnly);
arg << 5;
if (!client->send("someAppId", "fooObject/barObject", "doIt(int)",
	          data))
  tqDebug("there was some error using DCOP.");

OK, now let's say we wanted to get the data back from the remotely
called function.  You have to execute a call() instead of a send().
The returned value will then be available in the data parameter "reply".
The actual return value of call() is still whether or not DCOP
communication was successful.

QByteArray data, replyData;
QCString replyType;
QDataStream arg(data, IO_WriteOnly);
arg << 5;
if (!client->call("someAppId", "fooObject/barObject", "doIt(int)",
                  data, replyType, replyData))
  tqDebug("there was some error using DCOP.");
else {
  QDataStream reply(replyData, IO_ReadOnly);
  if (replyType == "QString") {
    QString result;
    reply >> result;
    print("the result is: %s",result.latin1());
  } else
    tqDebug("doIt returned an unexpected type of reply!");
}

N.B.: You cannot call() a method belonging to an application which has
registered with an unique numeric id appended to its textual name (see
dcopclient.h for more info). In this case, DCOP would not know which
application it should connect with to call the method. This is not an issue
with send(), as you can broadcast to all applications that have registered
with appname-<numeric_id> by using a wildcard (e.g. 'konsole-*'), which
will send your signal to all applications called 'konsole'.

Receiving Data via DCOP:
------------------------

Currently the only real way to receive data from DCOP is to multiply
inherit from the normal class that you are inheriting (usually some
sort of QWidget subclass or QObject) as well as the DCOPObject class.
DCOPObject provides one very important method: DCOPObject::process().
This is a pure virtual method that you must implement in order to
process DCOP messages that you receive.  It takes a function
signature, QByteArray of parameters, and a reference to a QByteArray
for the reply data that you must fill in.

Think of DCOPObject::process() as a sort of dispatch agent.  In the
future, there will probably be a precompiler for your sources to write
this method for you.  However, until that point you need to examine
the incoming function signature and take action accordingly.  Here is
an example implementation.

bool BarObject::process(const QCString &fun, const QByteArray &data,
		        QCString &replyType, QByteArray &replyData)
{
  if (fun == "doIt(int)") {
    QDataStream arg(data, IO_ReadOnly);
    int i; // parameter
    arg >> i;
    QString result = self->doIt (i);
    QDataStream reply(replyData, IO_WriteOnly);
    reply << result;
    replyType = "QString";
    return true;
  } else {
    tqDebug("unknown function call to BarObject::process()");
    return false;
  }
}

Receiving Calls and processing them:
------------------------------------

If your applications is able to process incoming function calls
right away the above code is all you need. When your application
needs to do more complex tasks you might want to do the processing
out of 'process' function call and send the result back later when
it becomes available.

For this you can ask your DCOPClient for a transactionId. You can
then return from the 'process' function and when the result is
available finish the transaction. In the mean time your application
can receive incoming DCOP function calls from other clients.

Such code could like this:

bool BarObject::process(const QCString &fun, const QByteArray &data,
		        QCString &, QByteArray &)
{
  if (fun == "doIt(int)") {
    QDataStream arg(data, IO_ReadOnly);
    int i; // parameter
    arg >> i;
    QString result = self->doIt(i);

    DCOPClientTransaction *myTransaction;
    myTransaction = kapp->dcopClient()->beginTransaction();

    // start processing...
    // Calls slotProcessingDone when finished.
    startProcessing( myTransaction, i);

    return true;
  } else {
    tqDebug("unknown function call to BarObject::process()");
    return false;
  }
}

slotProcessingDone(DCOPClientTransaction *myTransaction, const QString &result)
{
    QCString replyType = "QString";
    QByteArray replyData;
    QDataStream reply(replyData, IO_WriteOnly);
    reply << result;
    kapp->dcopClient()->endTransaction( myTransaction, replyType, replyData );
}

DCOP Signals
------------

Sometimes a component wants to send notifications via DCOP to other
components but does not know which components will be interested in these
notifications. One could use a broadcast in such a case but this is a very
crude method. For a more sophisticated method DCOP signals have been invented.

DCOP signals are very similair to Qt signals, there are some differences 
though. A DCOP signal can be connected to a DCOP function. Whenever the DCOP
signal gets emitted, the DCOP functions to which the signal is connected are
being called. DCOP signals are, just like Qt signals, one way. They do not
provide a return value. 

A DCOP signal originates from a DCOP Object/DCOP Client combination (sender). 
It can be connected to a function of another DCOP Object/DCOP Client 
combination (receiver).

There are two major differences between connections of Qt signals and 
connections of DCOP signals. In DCOP, unlike Qt, a signal connections can
have an anonymous sender and, unlike Qt, a DCOP signal connection can be
non-volatile.

With DCOP one can connect a signal without specifying the sending DCOP Object 
or DCOP Client. In that case signals from any DCOP Object and/or DCOP Client
will be delivered. This allows the specification of certain events without
tying oneself to a certain object that implementes the events.

Another DCOP feature are so called non-volatile connections. With Qt signal
connections, the connection gets deleted when either sender or receiver of
the signal gets deleted. A volatile DCOP signal connection will behave the
same. However, a non-volatile DCOP signal connection will not get deleted 
when the sending object gets deleted. Once a new object gets created with 
the same name as the original sending object, the connection will be restored.
There is no difference between the two when the receiving object gets deleted,
in that case the signal connection will always be deleted.

A receiver can create a non-volatile connection while the sender doesn't (yet)
exist. An anonymous DCOP connection should always be non-volatile.

The following example shows how KLauncher emits a signal whenever it notices
that an application that was started via KLauncher terminates.

   QByteArray params;
   QDataStream stream(params, IO_WriteOnly);
   stream << pid;
   kapp->dcopClient()->emitDCOPSignal("clientDied(pid_t)", params);

The task manager of the KDE panel connects to this signal. It uses an 
anonymous connection (it doesn't require that the signal is being emitted
by KLauncher) that is non-volatile:

   connectDCOPSignal(0, 0, "clientDied(pid_t)", "clientDied(pid_t)", false);

It connects the clientDied(pid_t) signal to its own clientDied(pid_t) DCOP
function. In this case the signal and the function to call have the same name.
This isn't needed as long as the arguments of both signal and receiving function
match. The receiving function may ignore one or more of the trailing arguments
of the signal. E.g. it is allowed to connect the clientDied(pid_t) signal to
a clientDied(void) DCOP function.

Using the dcopidl compiler
---------------------

dcopidl makes setting up a DCOP server easy. Instead of having to implement
the process() method and unmarshalling (retrieving from QByteArray) parameters
manually, you can let dcopidl create the necessary code on your behalf.

This also allows you to describe the interface for your class in a
single, separate header file.

Writing an IDL file is very similar to writing a normal C++ header. An
exception is the keyword 'ASYNC'. It indicates that a call to this
function shall be processed asynchronously. For the C++ compiler, it
expands to 'void'.

Example:

#ifndef MY_INTERFACE_H
#define MY_INTERFACE_H

#include <dcopobject.h>

class MyInterface : virtual public DCOPObject
{
  K_DCOP

  k_dcop:

    virtual ASYNC myAsynchronousMethod(QString someParameter) = 0;
    virtual QRect mySynchronousMethod() = 0;
};

#endif

As you can see, you're essentially declaring an abstract base class, which
virtually inherits from DCOPObject.

If you're using the standard KDE build scripts, then you can simply
add this file (which you would call MyInterface.h) to your sources
directory. Then you edit your Makefile.am, adding 'MyInterface.skel'
to your SOURCES list and MyInterface.h to include_HEADERS.

The build scripts will use dcopidl to parse MyInterface.h, converting
it to an XML description in MyInterface.kidl. Next, a file called
MyInterface_skel.cpp will automatically be created, compiled and
linked with your binary.

The next thing you have to do is to choose which of your classes will
implement the interface described in MyInterface.h. Alter the inheritance
of this class such that it virtually inherits from MyInterface. Then
add declarations to your class interface similar to those on MyInterface.h,
but virtual, not pure virtual.

Example:

class MyClass: public QObject, virtual public MyInterface
{
  Q_OBJECT
  
  public:
    MyClass(); 
    ~MyClass();

    ASYNC myAsynchronousMethod(QString someParameter);
    QRect mySynchronousMethod();
};

Note: (Qt issue) Remember that if you are inheriting from QObject, you must
place it first in the list of inherited classes.

In the implementation of your class' ctor, you must explicitly initialize
those classes from which you are inheriting from. This is, of course, good
practice, but it is essential here as you need to tell DCOPObject the name of
the interface which your are implementing.

Example:

MyClass::MyClass()
  : QObject(),
    DCOPObject("MyInterface")
{
  // whatever...
}

Now you can simply implement the methods you have declared in your interface,
exactly the same as you would normally.

Example:

void MyClass::myAsynchronousMethod(QString someParameter)
{
  tqDebug("myAsyncMethod called with param `" + someParameter + "'");
}


It is not necessary (though very clean) to define an interface as an
abstract class of its own, like we did in the example above. We could
just as well have defined a k_dcop section directly within MyClass:

class MyClass: public QObject, virtual public DCOPObject
{
  Q_OBJECT
  K_DCOP

  public:
    MyClass(); 
    ~MyClass();

  k_dcop:
    ASYNC myAsynchronousMethod(QString someParameter);
    QRect mySynchronousMethod();
};

In addition to skeletons, dcopidl2cpp also generate stubs. Those make
it easy to call a DCOP interface without doing the marshalling
manually. To use a stub, add MyInterface.stub to the SOURCES list of
your Makefile.am. The stub class will then be called MyInterface_stub.

Conclusion:
-----------

Hopefully this document will get you well on your way into the world
of inter-process communication with KDE!  Please direct all comments
and/or suggestions to Preston Brown <[email protected]> and Matthias
Ettrich <[email protected]>.


Inter-user communication
------------------------

Sometimes it might be interesting to use DCOP between processes
belonging to different users, e.g. a frontend process running
with the user's id, and a backend process running as root.

For this you can use tdesu with the --nonewdcop option. tdesu will
then forward the address of the DCOP server as well as the authentication
information to the new user. 

*WARNING*: This gives the user that you su to, full access to your session!
If you su to root this will not be a problem, but it may be a problem if
you su to another user.

By default, KDE applications (e.g. the ones that run as root) that connect
to the dcopserver of another user will not accept any incoming DCOP calls.
You can override this with DCOPClient::setAcceptCalls() after you have
carefully reviewed that your DCOPClient does not provide objects/functions
that could be abused for privilege escalation.


Example: tdesu --nonewdcop -u root -c kcmroot

This will, after tdesu got the root password, execute kcmroot as root,
talking to the user's dcop server.


Performance Tests:
------------------
A few back-of-the-napkin tests folks:

Code:

#include <kapplication.h>

int main(int argc, char **argv)
{
  KApplication *app;

  app = new KApplication(argc, argv, "testit");
  return app->exec();
}

Compiled with:

g++ -O2 -o testit testit.cpp -I$QTDIR/include -L$QTDIR/lib -ltdecore

on Linux yields the following memory use statistics:

VmSize:     8076 kB
VmLck:         0 kB
VmRSS:      4532 kB
VmData:      208 kB
VmStk:        20 kB
VmExe:         4 kB
VmLib:      6588 kB

If I create the KApplication's DCOPClient, and call attach() and
registerAs(), it changes to this:

VmSize:     8080 kB
VmLck:         0 kB
VmRSS:      4624 kB
VmData:      208 kB
VmStk:        20 kB
VmExe:         4 kB
VmLib:      6588 kB

Basically it appears that using DCOP causes 100k more memory to be
resident, but no more data or stack.  So this will be shared between all
processes, right?  100k to enable DCOP in all apps doesn't seem bad at
all. :)

OK now for some timings.  Just creating a KApplication and then exiting
(i.e. removing the call to KApplication::exec) takes this much time:

0.28user 0.02system 0:00.32elapsed 92%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1084major+62minor)pagefaults 0swaps

I.e. about 1/3 of a second on my PII-233.  Now, if we create our DCOP
object and attach to the server, it takes this long:

0.27user 0.03system 0:00.34elapsed 87%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1107major+65minor)pagefaults 0swaps

I.e. about 1/3 of a second.  Basically DCOPClient creation and attaching
gets lost in the statistical variation ("noise").  I was getting times
between .32 and .48 over several runs for both of the example programs, so
obviously system load is more relevant than the extra two calls to
DCOPClient::attach and DCOPClient::registerAs, as well as the actual
DCOPClient constructor time.