Ganymede 0.99.7 released

Jonathan Abbey jonabbey at arlut.utexas.edu
Wed Oct 13 21:26:48 UTC 1999


Ganymede 0.99.7 is available at

 ftp://ftp.arlut.utexas.edu/pub/ganymede/

 or

 ftp://129.116.212.1/pub/ganymede/

if your client has trouble resolving a proper address for
ftp.arlut.utexas.edu.

Ganymede is a GPL'ed system for managing NIS, DNS, and other
network directory services in medium-to-large environments.

The Ganymede web page is at http://www.arlut.utexas.edu/gash2/.

--

It's been almost 6 weeks since the last Ganymede release, and lots of
things have been fixed and improved.  We went into production with
Ganymede this past week, and lots of the new features and fixes came
about as part of that.

Highlights include a fix for the old 'can't bind to localhost' problem
that people had been reporting when running the Ganymede server on
Linux.  Thanks to Michael McEniry (mmceniry at itsc.uah.edu) for
providing a very nice patch to the Ganymede server to take care of
this problem.

I had to take care of a whole cluster of problems from 0.99.6 that
made it basically impossible to properly take a newly dumped schema
file and load data into it using any of the included schema kits.  I
also fixed a bug introduced in 0.99.6 that broke the database rollback
code.

The new features include some extra sanity checking in the server's
start-up code, a reworking of the builder task classes to not leak
file descriptors, an idle user time-out server thread, the ability for
the client to reconnect to the Ganymede server if the server is
restarted while the client is loaded, and a deferred shutdown option
for the server, to cleanly shut down the system without having users
lose changes they are working on.

-------------------- Changes from 0.99.6 to 0.99.7 -------------------

RELEASE DATE: October 13, 1999

1. [SERVER] Fixed the Ganymede server's registry binding for localhost

Michael McEniry (mmceniry at itsc.uah.edu) provided a patch to
Ganymede.java which corrects the problem on Linux systems with the
server failing on startup with a localhost RMI registry binding
permissions failure.

2. [SERVER] Fixed object deletion for multi-level nested objects

Martin Schneider (martin at ira.uka.de) wrote to report that deleting
objects which contained embedded objects which contain embedded
objects fails on transaction commit.. the problem was that the
GanymedeSession.getContainingObj() method couldn't walk the embedding
stack to find the top-level parent to determine object ownership for
logging purposes.

GanymedeSession.getContainingObj() will now check to see if it is
trying to find the parent of an embedded object which has been
processed for deletion, in which case it will check the original
version of the object in order to find the original parent.

3. [SERVER] Added support for regexp matching in StringDBField

Added support for specifying a regular expression pattern in string
fields.  Modified the server and schema editor to support regular
expression binding to string fields.

The older 'OkChars' and 'BadChars' string field parameters are still
maintained.. any change to a string field must satisfy any regular
expression specified in addition to the OkChars and BadChars
parameters.

This change bumps the DBStore revision numbers to

	major: 1
	minor: 14

Martin Schneider (martin at ira.uka.de) asked for this change, it seemed
like a good idea to me.

4. [SERVER] Fixed server-loading bug from 0.99.6.6

Change 6 to Ganymede.java in release 0.99.6 broke loading data from a
freshly dumped schema.  Fixed the server initialization so that it can
properly reset passwords if the password field is undefined in the
supergash persona object.

5. [SERVER] Fixed schema dump bugs in DBObject.partialEmit()

Since July there has been a bug in DBObject.partialEmit() which makes
it impossible to successfully dump a modified schema to disk and then
turn around and use it to load a new database.  The problem was that
when the server was dumping a ganymede.schema file, the supergash
persona was being written to disk without the link to the supergash
owner group (owner group 0) that gave it supergash privileges.

Actually, there were a cluster of bugs in the schema dump logic.  All
of them have been fixed, so schema dumping should produce a clean and
usable ganymede.schema file now.  I knew the code for schema dumping
was poorly architected, but I somehow missed the fact that it was so
thoroughly broken.

6. [SERVER] Added code to protect the essential server data structures

It used to be possible to manually delete the supergash owner group,
which would put the ganymede.db file in an unusable state.
GanymedeSession.remove_db_object() now refuses attempts to do this.

In addition, the Ganymede server now checks the database schema on
start-up to make sure that all of the essential database objects are
defined, including the supergash persona object, the supergash owner
group, the default role, and the system event objects.  If any of
these objects are undefined in the database on server startup, the
server will now recreate them.

The server still doesn't do a schema sanity check on startup, though.
Still lots more work to do to clean up schema editing as a whole.

7. [SERVER] Fixed logging for IP address changes

The IPDBField.getDiffString() method had an error that caused it to
report the old value for a changed IP address when it was meant to be
showing the new.

8. [SERVER] Added a background task to disconnect idle users

The Ganymede server now has a built-in task which runs every minute to
disconnect users that have been idle for a certain period of time.

Currently, users with no objects checked out of the database will be
disconnected after 15 minutes of idle time, and users who are in the
middle of a transaction but have not done anything for a long time
will be disconnected after 20 minutes of idle time.

9. [CLIENT, SERVER] Made client able to re-acquire server without restarting

The Ganymede client is now able to attempt to re-acquire the Ganymede
server if the server goes down while the client is running.

That is, if the server goes down cleanly while the client is running,
the client will close the main window, leaving the login box.  The
login box will now be able to attempt to bind to a new instantiation
of the Ganymede server if the server has been brought back up.

In order to do this properly, I had to add a new method to the
RMI Server interface (Server.up()), so you'll need to install
the client from Ganymede 0.99.7 for use with the 0.99.7 server,
and vice-versa.

Also modified the client to be more responsive to a server shut-down
message, even if the user was sitting on the persona selection dialog.

10. [SERVER] Server now associates embedded objects with container in log

Previously, if changes were made to an embedded object that did not
result in a modification to the state of the container, the log for
that change did not record that the container was involved in the
transaction, making such changes invisible to the history panel of the
containing object.

Now, whenever an embedded object is changed, the log record for the
change will be associated its container and the change will show up
properly in the container's history view.

11. [SERVER,SCHEMA,DISTRIBUTION] Changes to handle descriptor starvation

We put Ganymede in production at the laboratory on October 10th and
quickly discovered that the Ganymede server was running into file
descriptor exhaustion when builds are done at a high rate of speed.

It turns out that one of the problems was that the GASHBuilderTask
class in the gasharl schema kit wasn't explicitly closing the file
descriptors for the Process object used to do the external build.
Looking on the Java BugParade, I found bug item #4025386, which
describes the problem and how to fix it.

I went through all the BuilderTask classes in all the schema kits
included with Ganymede and added a finally {} clause to close the
Process file descriptors once the external build was done.

Even after this change, I still saw the server run out of descriptors,
so I put a 'ulimit -n 128' in the runServer script (64 is the default
on Solaris.. the ulimit command is commented out so that people on
Linux, etc., don't get an artificially low descriptor limit), and I
fixed a few other places in the code where the system was
unnecessarily waiting for garbage collection to close resources.

File descriptor depletion is still a big issue with Ganymede, as a
user with the client login window open is taking up a descriptor on
the server for the RMI connection, even though the user isn't doing
anything.  I'm not sure how much I can do about this, since I believe
the only way to force an RMI break is by allowing the remote object to
be garbage collected on the client.

In any case, the problem isn't so much that the Ganymede server itself
is keeping an excessive number of descriptors open, but that the
extensive Ganymede/GASH build process we initiate on transaction
commit opens dozens of files, taking up lots of the server's
descriptors along the way.  Java doesn't provide any good way to
manage this, but it may be a good idea to put another ulimit
invocation in the script run by Ganymede?

12. [SERVER] Fixed really stupid DBEditSet rollback mistake

Ugh. Ugh. Ugh.

I had broken the DBEditSet rollback code with change #0.99.6.9 below.

There was no way it could have ever worked.  It didn't come close to
working.  Not even remotely plausible code.  I am dum.

I fixed it.. it is now both efficient and correct.

13. [SERVER] Optimized DBEditSet.commit()

I reworked DBEditSet.commit() so that it doesn't need to use
rollback() to handle an incompletely filled-out object at transaction
commit time.  Basically streamlined commit() so that it does the
minimum amount of work required to check things out, without doing an
unnecessary checkpoint()/rollback() pair when one or more objects in a
transaction aren't completely filled out.

And I checked my work, and I tested the code, this time.

14. [SERVER, ADMIN CONSOLE] Added a deferred shutdown option

The server can now be told to disable logins and to shutdown the server
when all users log out of the system.  This is very useful for shutting
the server down cleanly in a production environment.

If you use the 'stopServer' script, you can now say 'stopServer
-delay' to have the server do a user-friendly shutdown.  If you do do
'stopServer -delay', the stopServer script will not wait for the
server to shut down before exiting.

If you shut the server down through the admin console, a dialog will
pop up asking if you want to make the shut down deferred or not.

This change involves an incompatible change to the
arlut.csd.ganymede.adminSession RMI remote interface used by the
server, admin console, and stopServer script.

15. [SCHEMA] Several tweaks to gasharl schema

Lots of minor fix-ups to the gasharl schema code, which we are actually
using in production at ARL now.

-------------------------------------------------------------------------------
Jonathan Abbey 				              jonabbey at arlut.utexas.edu
Applied Research Laboratories                 The University of Texas at Austin
Ganymede, a free NIS/DNS management system    http://www.arlut.utexas.edu/gash2


More information about the bind-users mailing list