=============================================================================
# Number: 1
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 03.11.93
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by:
# j_nieplocha@pnlg.pnl.gov (Dr. Jaroslaw Nieplocha)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
Improper input argument screening in masktrap() and _set_trap_mask()

instead of:
   if ((state != 0) || (state != 1))
there should be:
   if ((state != 0) && (state != 1))

# Location/Fix: by: Lamberts
Applied proposed fix.

=============================================================================
# Number: 2
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 03.11.93
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by:
# j_nieplocha@pnlg.pnl.gov (Dr. Jaroslaw Nieplocha)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
hrecv seems not to be working -- the message is never received.

# Location/Fix: by: Lamberts

In

buffer/call_buffer.c(213) : ...(bce->el_type == HRECVX))
                                                ^^

had to be fixed.

=============================================================================
# Number: 3
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 29.12.93
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by:
# Stefan Lamberts
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
c/isend to node -1 does not work if the partition has size 1

# Location/Fix: by: Lamberts

The Problem is that __isend() returns a message id -1 if no message is
sent. Obviously this is not totaly wrong. A possible fix is to have a
dummy message id which doesn't belong to any real message. msgdone()
called with that message id should return ok.

What happens on the Paragon?

The Paragon allows it!

Fixed by Stefan Lamberts 18.01.94

Added procedures _dummy_msgid() and _is_dummy_msgid().
In _isend() a dummy message id is used for multiple sends with
nodesel == -1. The msg...() calls are changed in a way that they can
be called with the dummy id.


=============================================================================
# Number: 4
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
Message passing from the controlling process.

Does the address resolution work for the controlling process?

Report fixes to Chris Elford.

Fixed by Susanna Schink

Lambo

=============================================================================
# Number: 5
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
nx_load doen't work propperly on application processes. 

Chris Elford sent an example program that works on the Paragon.

More information on this form Chris Elford:

 On the nx_load() problem, I found the following:

In nxcalls/nxload_d.c _dd_pidnr() is forwarding a rel_send((long)DA_APEXIT
instead of a _rel_send((long)DA_PIDNR.  In addition, 

In reliable/a_handler.c, there was no support for an incoming DA_PIDNR.
I added 
  case DA_PIDNR:
    res = _da_pidnr(msgd);
    break;

LAMBO: Chris is right about that.

I found this because on nx_load, I was getting weird loaded process exited
messages.  These modifications allow the nx_load() to return in the caller.
Unfortunately, there is still a problem.  in nxcalls/nxload_d.c,
the call to forkexec(msgd) returns zero when not the original call.  This
traces down to the call to _ins_plistel() returning <0 at the end of forkexec().
_ins_plistel got too complicated for me to track into though.  The problem is 
that the addressing table is messed up.

LAMBO: This has to be done!!!

Hope this is helpful.
Chris (elford@cs.uiuc.edu)

Fixed by Susanna Schink

Lambo

=============================================================================
# Number: 6
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
_cprobex doesn't allow typesel == -1.

In nxcalls/probe.c in _cprobex() it should be

if (((typesel != -1) && (typesel < 0).....
      ^^^^^^^^^^^^^ this is missing.

Fixed by Stefan Lamberts 18.01.94

=============================================================================
# Number: 7
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
The array info isn't set propperly in _iprobex().

in nxcalls/probe.c it should be

info[3] = bme->s_ptype;
     ^ this should be 3 not 4.

Fixed by Stefan Lamberts 18.01.94

=============================================================================
# Number: 8
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:
setptype(0); setptype(0); setptype(0);
should work without returning an error.

Fix:
NxlibV1_0/nxcalls/ptype.c for _setptype()

  /* ptype allready set */
  if (val == ptype)
  {
    errno = EQSET;
    return (-1);
  }

becomes

  /* ptype allready set to this value */
  if (val == ptype)
    return (0);


Fixed by Stefan Lamberts 18.01.94


=============================================================================
# Number: 8
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 03.02.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Stefan Lamberts
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

The name of the file remote/remote_signal.c is to long for ar under
HP_UX and SGI_IRIX. This should be fixed.

Fixed by Lambo (11.03.94)

=============================================================================
# Number: 9
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 11.02.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Christopher W. S. Bruner (cwsb@hellcat.sunbelt.nawcad.navy.mil)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

gcolx() doesn't match the NX gcolx() call.

The syntax of the Paragon gcolx() call is:

void gcolx( char x[], long xlen[], char y[])

Implemented is the following:

void gcolx( char x[], long xlen, char y[])

Fixed by Stefan Lamberts (11.02.94)

=============================================================================
# Number: 10
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford (elford@cs.uiuc.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

Messages in the new Posix version get lost sometimes.

This happens because some SIGIOs get lost. I'm not sure what the
reason is. It might be a problem with the sigaction() call. A similar
problem occurred in MMK/X by using the signal() call. It didn't happen
in MMK/X with sigvec. 

SunOS4 version used BSD signal calls now.

Fixed by Stefan Lamberts

Lambo

*****************************************************************************
*****************************************************************************
Version 1.1
*****************************************************************************
*****************************************************************************

=============================================================================
# Number: 11
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Makan.Pourzandi@lip.ens-lyon.fr
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

gdsum doesn't work on some machines.

In the original implementation the working array was not used for
global operation. Instead the message with the gathered data was used.
That led to a alignment problem which produced a BUS ERROR and a core
dump.

The bug was fixed by using the working array. The data is now copied
to the work array which is used for the global operation.
Modifications were made in global/glocal_calls.c

Fixed by Stefan Lamberts (1994--03--30)

Lambo

=============================================================================
# Number: 12
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 17.01.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Lambo
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

setpgid() caused problems on RS6K if the nxdaemon is executed on a
remote machine. 

In daemon/nxdaemon.c setpgig() is no longer called on RS6Ks.

Fixed by Stefan Lamberts (1994-03-30)

Lambo

*****************************************************************************
Version 1.1.1
*****************************************************************************



=============================================================================
# Number: 13
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 07.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Lambo
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

The message about the exit status respectively the signal that
terminated the process should be printed to stderr instead of stdout.
Right now _dbgmsg() is used for this message. printf() to stderr seems
better.

Fixed by Stefan Lamberts 07.04.94

Lambo

=============================================================================
# Number: 14
# State : reported (reported, accertained, no bug, located, fixed)
# Date  : 07.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Lambo
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:


Hostnames in the partition file should be conform to RFC952 and
RFC1132. IP adresses should be supported as well.

A vaild hostname is the following regular expression:
[A-Za-z0-9][-.A-Za-z0-9]{0,253}[A-Za-z0-9]

An IP address is:
DEC[.]DEC[.]DEC[.]DEC

where DEC is [1-9][0-9]* | 0


Additionally it should be possible that pathnames in the partition
file contain environment variables. At the moment a pathname can only
be either an environment variable or a pathname. It should be
considered if and how environment variables on the remote should be
supported. Beware that a pathname may contain allmost any character.
This should be supported as well.

=============================================================================
# Number: 15
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 15.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Lambo
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

nx_initve_ in fortran/fortran.c line 1423 was misspelled.

Fixed by Lambo  1994-04-15


=============================================================================
# Number: 16
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 22.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

Finally malloc causes problems. all memory allocation calls used
within the application shlud be embraced with _rel_disable()
_rel_enable().

A macro in nx.h would do. All calls should be replaced by nxlib_...
calls which do just that.

A problem might be malloc calls within NXLib. If they are within the
interrupt handler they should not be embraced. otherwise they should
be.

The malloc calls within NXLib are now bracketted with
_rel_disable()/_rel_enable(). It was necessary to modify theese
procedures in order to do so. There are new procedures to determine
wether the processis within the signal handler.



=============================================================================
# Number: 17
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 22.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

_rel_disable() _rel_enable() produces warnings of mismatched pairs due
to NX calls within hrecv handlers.

This can be avoided by using a counter for _rel_disable _rel_enabe().

Now there is a counter for _rel_disable()/_rel_enable().

Fixed by Stefan Lamberts 94-05-02

  
=============================================================================
# Number: 18
# State : accertained (reported, accertained, no bug, located, fixed)
# Date  : 22.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Chris Elford
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

NXLib does not allow for reusing an old ptype even if the process that
had the ptype before terminated. On a Paragon this is possible. 

To fix this bug the daemon process must update it's data base AND the
data bases of other processes. Connections wich are already
established have to be reopend if needed. The biggest problem is to
let the other daemon processes now of the new address.

=============================================================================
# Number: 19
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 26.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Achim Jung
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

mallopt() expects two arguments. Only one is passed to it.

Fixed by Stefan Lamberts 94-04-26

=============================================================================
# Number: 20
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 26.04.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Tobias Fischer
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

In _conf_calls() (file buffer/call_buffer.c it was not checked wether
the call that will be confirmed was a receive call. Pending isend()
calls were confirmed as well. That was nonsense. 

Fixed by Stefan Lamberts 94-04-29

*****************************************************************************
Version V1_1_2
*****************************************************************************

=============================================================================
# Number: 21
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 18.07.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# fyodor@nas.nasa.gov (Chris L. Kuszmaul)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

 In nxinit.c, w4e have:

   if ((pid < (pid_t)0) && (errno = ECHILD))

  AT or near line 146. I suspect the = should be ==.

  CLK


This is obviously a bug

Fixed by Stefan Lamberts 94-07-18

=============================================================================
# Number: 22
# State : acertained, located, workaround (reported, accertained, no bug, located, fixed)
# Date  : 18.07.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Stefan Lamberts
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

In rare cases if processes exchange messages vica versa it might
happen that theese processes deadlock.
That happens if both processes start writing the data simultaneously
and are interrupted by a SIGIO signal before completing the write call
by the data comming from the other process. Then, they both atempt to
read the complete message from the other process which is obviously
not available.

It would require quite a lot of work to fix this bug. It would be
necessary to keep messages form sockets in mind which were not
complete and how many data needs to be read additionally. Then a
blocking read call could abort and the write call could continue.
Another possibility is to make TCP handle the problem. If TCP is able
to buffer the complete message, just write it without beeing
interrupted. 

Right now we decided NOT to fix this bug. Nobody in the users
community complained about it so that we see no need for a fix.

We implemented a kind of workaround. This is done by increasing the
TCP send and receive buffer sizes. Due to this fix the application now
understands an additional commandline option -tcpbuf which is used to
set the TCP send and receive buffers to the given value.

Stefan Lamberts 94-08-17


=============================================================================
# Number: 23
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 16.08.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Phillip Dickens <pmd@icase.edu>
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:


In file addr_db.c:

When the maximum file descriptor numer is set, the array of
descriptors rcdarr is adressed form rcdsiz to 0. This is obviously a
bug since rcdarr has only rcdsiz fields. The index should start at
rcdsiz-1. 

Fixed by Stefan Lamberts 94-08-16


=============================================================================
# Number: 24
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 08.08.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Phil ??? (phil@ctrhdc2.ct.ornl.gov)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

When using wildcard processor numbers in NX_INITVE or NX_LOAD with
fortran it causes problems on some systems.

In some of the fortran stubs arrays are allocated with malloc(). If a
wildcard is used, a negative number of bytes is allocated. This causes
an error on some systems.

Fixed by Stefan Lamberts 94-08-10

=============================================================================
# Number: 25
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 29.08.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Phil ???  (phil@ctrhdc2.ct.ornl.gov)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

[ic]send() to the same process doesn't work.

The problem is that the the confirmation message was handled
incorrectly. Additionally messages must be buffered in a new memory
area if there is no matching receive call pending.

Fixed by Stefan Lamberts 94-08-29

*****************************************************************************
Version V1_1_3
*****************************************************************************


=============================================================================
# Number: 26
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 02.09.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# (Wang Dao-liu (wdl@indigo9.cc.ac.cn)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

In the RSHPATH macro for SGI a leading '/' was missing.

Fixed by Stefan Lamberts 94-09-02


=============================================================================
# Number: 27
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 16.11.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Geoffrey Furnish (furnish@dino.ph.utexas.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

I just noticed that your nx.h has 
        long msgwait()
whereas the doc (page 14, A.2.3) has it as
        void msgwait ...

void is correct.

Fixed by Stefan Lamberts 94-11-16

=============================================================================
# Number: 28
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 21.12.94
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Geoffrey Furnish (furnish@dino.ph.utexas.edu)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

In a configuration of bursty message transfers between many nodes the
nxdaemon dumps a core after a segmantation violation.

This happend because the pointer to the next element in the list of
unanswered address request was not initialized. In proceudre ins_req()
the initialization of re->next = (req_el *)0 had to added.

Fixed by Stefan Lamberts 94-12-21

=============================================================================
# Number: 29
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 06.02.95
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Henning Otto (h.otto@kfa-juelich.de)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:


Ihr Message-Passing-Simulator verhaelt sich in dabei leider
etwas anders als die Paragon es tut. Setzt man innerhalb
einer Message-Behandlungsroutine erneut ein weiteres hrecv auf,
so wird manchmal die neue Behandlungsroutine unmittelbar
aktiviert, auch wenn die entsprechende Nachricht noch nicht
eingetroffen ist. Dieses Verhalten zeigt sich sowohl unter
SUN-OS als auch unter SOLARIS.

The problem was, that hcalls were not removed from the list of calls
which need to be handled before the handler was called.

Fixed by Stefan Lamberts 95-02-06

=============================================================================
# Number: 30
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 06.02.95
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# Nicolas Droux <droux@info.isbiel.ch>
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

I'm trying to run one of my Paragon programs on NXLib and I have a problem with 
 
standard IO. One of the nodes asks the user to give some parameters on
startup with a call to gets() and sscanf(). Unfortunately, although
the messages from the nodes are printed on the terminal of the user,
the gets() or scanf() functions never return. Note that only one of
the node interact directly with the user.

Is this a limitation of NXLib ? Or is there any well-known workaround
for this problem ? I understand that this is a difficult problem to
handle in distributed and parallel system, but I really need to be
able to use stdin from one of the nodes.


This is a known problem. The reason is that processes in the backgroud
cannot read form stdin. Application processes running on remote
machines cannot read from stdin in any case. I wrote a shell skipt
which enables reading form stdin for local processes. If more thatn
one process reads from stdin it is undefined whether one or more
processes will read and which of the processes will get the input.


=============================================================================
# Number: 31
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 13.02.95
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# R.Berrendorf (r.berrendorf@kfa-juelich.de)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

Sending or receiving messages with 0 length from or to a buffer
(char *)0 returned an error.

Fixed by Stefan Lamberts 95-02-13

=============================================================================
# Number: 32
# State : fixed (reported, accertained, no bug, located, fixed)
# Date  : 13.02.95
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Reported by: 
# R.Berrendorf (r.berrendorf@kfa-juelich.de)
# 
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Description:

Es gibt Probleme mit dem Message Passing, wenn man Broadcast macht und
mehr Knoten als physikalische Prozessoren vorhanden sind. Hier ein
kurzes Beispielprogramm, das auf 2 Knoten auf der 2-Prozessor-Sun
laeuft, aber auf 3 Knoten haengen bleibt (2 Knoten terminieren, der
dritte (sendende Knoten 0) bleibt haengen). Wenn ich explizit die
Message an Knoten 1 und 2 schicke, laeuft das Programm.

In _msgdone() was a bug if merged messages were free()d.

Fixed by Stefan Lamberts 95-02-14

