Here's another point to consider.

I just submitted a bug report concerning serialization of custom blocks (I
assume marshalling uses the serialization code). The deserialization code
for custom blocks does not correctly reproduce sharing. Try writing a tuple
of the same int64 data structure and reading it back via output_value,
input_value. The second component of the built tuple is garbage.

I submitted a fix which works for me:

Add:

At line 266 in file byterun/intern.c, insert the following line:

        if (intern_obj_table != NULL) intern_obj_table[obj_counter++] = v;


There's a int64 related bug in byterun/ints.c.

Change line:282 of byterun/ints.c
  <
  *wsize_64 = *wsize_64 = 8;
  >
  *wsize_32 = *wsize_64 = 8;

This bug caused the size of the block to be allocated when reading int64
data back in to be wrong on 32bit architectures.

-Manuel


-----Original Message-----
From: Gerd Stolpmann [mailto:gerd@gerd-stolpmann.de]
Sent: Tuesday, November 28, 2000 5:22 PM
To: Ken Wakita; jserot@lasmea.univ-bpclermont.fr
Cc: caml-list@inria.fr
Subject: Re: Marshaling of custom blocks and threads


On Tue, 28 Nov 2000, Ken Wakita wrote:
>I think the Marshal module is thread unsafe because it uses a shared
>buffer to produce the external image of the ML objects.

In 2.04, this was definitely true
(see http://caml.inria.fr/bin/caml-bugs/fixed?id=24;page=1;user=guest
and http://caml.inria.fr/bin/caml-bugs/fixed?id=25page=1;user=guest).
However, the bug was still subtle; the shared buffers were relatively 
well protected by the so-called "master lock". Until now, I thought
that 3.00 fixed the bug finally; i.e. the master lock really locked
the shared buffer. (I checked the code.)

Jocelyn Serot reports that Marshal is not thread-safe for custom blocks.
These did not exist in 2.04, so I suppose that only the new parts of the
marshalling system do not work properly enough. However, multi-threading
programming is error-prone, so it would be helpful to have a small program
that
demonstrates the bug (normally it is sufficient to repeat the errorneous
piece
of code often enough (e.g. 10000 times) to reproduce the incorrect
behaviour).

Gerd
-- 
----------------------------------------------------------------------------
Gerd Stolpmann      Telefon: +49 6151 997705 (privat)
Viktoriastr. 100             
64293 Darmstadt     EMail:   gerd@gerd-stolpmann.de
Germany                     
----------------------------------------------------------------------------