Phantasmal Site > DGD > The Kernel Library > Inheritance in the Kernel

Inheritance in the Kernel MUDLib

The main limitation the Kernel MUDLib imposes on inheritance is the division of objects into inheritables, clonables, and objects from which LWOs can be created

They're primarily distinguished from each other by what directory the files are in -- if the path contains a "/lib/" then the program is an inheritable. Inheritables may not be cloned, but may be inherited from other programs. They can't store any data, you can't find them with find_object(), and no functions (including create()) are ever called on them. The Kernel Library specifically prevents you from touching the 'real' object for reasons that will be explained later.

If a path contains a "/data/" then the object is a master for LightWeight Objects (LWOs) and can be instantiated with new_object(), but may not be cloned or inherited from. Note that a file with "/data/" in the path may inherit from inheritables just like everybody else. It also has data, and you can call functions on it. You can also call functions on LWOs that you make from it, and (of course) they can have their own copies of any data.

An object whose path contains "/obj/" is a cloneable. You can use clone_object() to make clones of it, but other programs can't inherit from it and you can't call new_object() on it to make an LWO from it. You can call functions on it (the master object and any of the clones), and it has usable data.

Any other path (one without either "/data/" or "/lib/" in the path) is assumed to be nothing special. You can't clone it, you can't make an LWO from it, you can't inherit from it. You can find it (remember, there's only the one master object) with find_object(), you can call functions on it and it can have data. Most daemons are this way in Kernel-derived MUDLibs. Usually authors will put "/sys/" in the path instead of "/obj/" or "/lib/" as documentation... But it's not required.

Okay, so why can't inheritables have data, or be cloned? Seems like a pretty serious limitation if you think about it a bit. It means that every parent class is an abstract parent class rather than being instantiable (to use some fancy Object Oriented terminology). The answer has everything to do with the way DGD allows you to recompile everything on the fly.

You can recompile an object with clones and at the end of that thread, all the clones get upgraded. It works for LWOs, too. That's pretty cool. You can recompile a library, and from then on any new objects that inherit it get the new version. Also cool. Unfortunately, old objects can't just switch which version of the parent class's code they use, so they're stuck with the old version of the library until you recompile them (after destructing and recompiling the library).

In the Kernel MUDLib you can deal with that -- just destruct the old version and recompile the clonable (see Issues, below, for an example). That'll upgrade all the clones, give you the library, and everything stays copacetic. Since a library has no clones and no data, when you destruct it and recompile, you lose nothing. Since the clonable has nothing inherit from it, you don't need to destruct it and recompile for the benefit of its child classes (since it has none).

So what if you don't do that? Melville and 2.4.5 both get away without doing any of this. The answer is that you can't upgrade some objects when the MUD is running, so you can kiss full-on persistence goodbye. To understand why, think about what an object which is both inheritable and clonable would be like to upgrade. In OO-speak, that's a concrete (non-abstract) parent class.

If the object were both inheritable and clonable then it could have some objects that inherited from it, plus a bunch of clones. If you wanted to upgrade it, you'd need to be able to destruct it and compile a new one so that the classes that inherit from it could get a new version. But that's a problem... You'd have to lose all the data in all the clones when you destructed it! So it can be both upgradable and clonable, but you couldn't upgrade its functions in its child classes without getting rid of all the clones... You can see why the Kernel Library just separates inheritables and clonables.

Could you make a different tradeoff to avoid having both child objects and clones? Probably. For instance, you could make an object be both inheritable and clonable, but make it destroy all its clones to upgrade. That'd be a massive pain, but you could do it. There are a lot of compromises like that, but the Kernel MUDLib's is simple, reliable and it works. If you want a different one, you can write a different MUDLib (or modify the Kernel MUDLib) to show us all how much better the world would be with your version...

With great power comes great responsibility, to quote an old comic book character you've all heard of. So the Kernel MUDLib makes a tradeoff -- you can recompile everything in the MUD if you want to, but in return you have this set of restrictions.

Recompilation and Object Issues

Each time an object is destructed and then compiled again from scratch, a new copy of it shows up. The old copy will silently go away when nobody's using it any more. This means that old versions of objects can float around for a very, very long time (remember that "Persistent MUD" idea?) in some cases.

How can you tell how it works? Well, when no existing object uses an old issue, it goes away. This uses Reference Counting, so when the reference count drops to zero, DGD knows that nobody's using it. So it has to have no clones (if clonable) and nobody can inherit from it (if inheritable). In either case, it must also have been destructed before it can go away from lack of references.

Here's an example:

Say A inherits from B, and B inherits from C...

  C <- B <- A
      

If I destruct C and recompile it, then B is out of date with C. It's using a previous issue of C. C now has two issues, the old and the new. If I destruct B and recompile it then the old B still inherits the old C. But the new B inherits the new C. So:

  oldC <- OldB <- A

  newC <- newB

If I then recompile A, that means nobody uses the old B or old C any more. Since I destructed them and nobody's using them, they'll finally go away. I would then only have one issue of B and one issue of C.

To spell that out further, let's arbitrarily assign some instance numbers to the issues.

Say the old A is issue #1, old B is #2, old C is #3. So,

  C(#3) <- B(#2) <- A(#1)
      

Now you recompile C (#3). So we have

  C(#3) <- B(#2) <- A(#1)

  C(#4)
      

That extra issue of C is just sitting off by itself. Nobody inherits from it. Then you destruct and recompile B:

  C(#3) <- B(#2) <- A(#1)

  C(#4) <- B(#5)
      

When B is recompiled, it looks up C to inherit from it. Issue #4 is the current non-destroyed one, so it finds that instead of #3. Then, if you recompile A in-place (for instance, if A is clonable so you don't *want* to destroy it):

  C(#4) <- B(#5) <- A(#1)
      

This assumes there's no other objects that reference the old B or C. If there aren't, then recompiling A (so it looks up B again) gets rid of the last reference to the old B(#2), which is destroyed. That removes the last ref to old C(#3), which is also destroyed. A is now linked to the new ones. If there are other objects that reference the old B(#2) or C(#3), then A will still be recompiled as above, but the old B and C will stick around longer, destructed but active.

Note that since A is recompiled in-place (instead of destructed and compiled), its issue number stays the same. You can find out the issue number in the Kernel MUDLib using either:

  status(obj)[O_INDEX];

or

  status(path)[O_INDEX];
      

In the first version, status() takes an object pointer which is for a clonable. The Kernel MUDLib will never give you an object pointer to an inheritable, so the second version takes the path string for the inheritable. I don't think the second version can look up destructed objects, only current non-destructed ones. If you write an object manager, you have to take that into account and keep track.

The Recompile Function

One thing that may confuse you is the recompile() function in the driver. Here's an explanation by Erwin Harte of when it would be called:

Date: Thu, 15 Feb 2001 17:26:32 +0100 (CET)
From: "Felix A. Croes"
Subject: Re: [DGD]kernel cloning and inheriting

Stephen Schmidt wrote:

>[...]
> Let me see if I have this straight; I'm pretty sure at
> some level I don't. Consider three objects, A, Ac, and B.
> Ac is a clone of A; B inherits A.

In the kernel library, inheritables are not clonable, but I
assume that you are talking about the general case.


> 1. You can recompile B because nothing inherits it.
> 2. You cannot recompile Ac because it is a clone; to recompile
>    Ac is basically to recompile A.

Yes.  Recompiling Ac is impossible, but if you could recompile A,
all clones would be upgraded as well.


> 3. You cannot recompile A because B inherits it.
> 4. You could recompile A if you first destructed B, but then
>    object B would be lost. In a persistant world, the loss of
>    B during the recompilation of A could be problematic.
> 5. You could recompile both B and A if Ac did not exist. (?)
>    The root problem is that if you try to recompile A in
>    the presence of Ac, then Ac is forced to change along
>    with B, and Ac might not want to do that.

If Ac did not exist, and A was a pure inheritable object, you could
have upgraded A and B with the following sequence:

    destruct A
    recompile B (will automatically compile A also)

As it is, you can still do this, but of course the state of A will
be lost thereby.  If A was a pure inheritable/clonable without its
own state, you can do this without negative effects on A and B.
Ac, however, will continue to use the old program of A, since A
was destructed before it was recompiled.  Only if A had been
recompiled without deing destructed first -- impossible because of
B -- would ac have been upgraded also.

Regards,
Dworkin

Date: Thu, 15 Feb 2001 18:47:04 +0100 (CET)
From: "Felix A. Croes"
Subject: Re: [DGD]kernel cloning and inheriting

Stephen Schmidt wrote:

>[...]
> > If Ac did not exist, and A was a pure inheritable object, you could
> > have upgraded A and B with the following sequence:
> > 
> >     destruct A
> >     recompile B (will automatically compile A also)
>
> If Ac did not exist, and you recompiled B without first
> destructing A, then A would not be recompiled? I'm pretty
> sure that's right.

Normally, it wouldn't be.  However, A may be out of date as well, in
the sense that it inherits yet another object -- let's call it C --
which has been destructed.  In that case, the recompilation of B
could in itself trigger the destruction of A from the recompile()
function in the driver object; the immediately following recompilation
of B would then also trigger the recompilation of A.


>[...]
> Just out of pure curiosity, why is it not possible to
> have changes in A reflected in both Ac and B? Is it
> because A doesn't keep track of a list of all objects
> that inherit it, so it doesn't know that when it updates
> its own code, it has to update B also? Or is it that when
> A is recompiled, you want B to keep the old behavior? Or
> is there something deeper going on?

Suppose that A is inherited by a further N objects.  If we want
to recompile A, and the changes must be automatically reflected
in those other, inheriting objects, than all of those objects
will have to be automatically upgraded.  I could have made it
that way, and in fact I originally intended to do so.  However,
if N becomes large, recompiling A is going to take a <very> long
time -- minutes in some existing DGD mudlibs.  I wanted to be
able to upgrade A without such a huge delay, so I decided to
divide the upgrade process into many smaller steps, each of which
could be done from LPC, perhaps with a series of callouts.  Since
this was completely new functionality and there was no backward
compatibility to take into account, I decided on the current
limitations.

Regards,
Dworkin