need for release?

[Filip Konvička]

Hi Boriss and all,

I have now a work-in-progress bugfix regarding pickles on Windows.

The problem is that the mechanism used for distinguishing between file 
and socket descriptors is done using the FD_XXX macros. This is bad, 
because the mechanism is not robust enough. Recently we are facing 
problems on long-running server machines, because the file descriptor 
handles are mistakenly used as socket descriptors, thus "send" is used 
on them instead of "write" (or vice-versa), resulting in errors.

Of course, this bug also affects many other things than pickles - the 
sockets and file io themselves, and many unexpected 
distributed-programming failures.

The solution is obvious - use a dynamic-sized unordered set for the 
socket descriptors. But - I need to know what are the choices for such 
implementation. Of course I don't want to introduce problems by writing 
my own hashtable implementation, so I'd like to use boost::unordered_set 
or at least std::map (this would degrade performance a bit in heavy-IO 
programs). But STD Lib is not used anywhere in Mozart, neither is boost, 
so I'm reluctant to introducing a new dependency.

I think that this is a critical bug and I'd be glad if we could fix it 
before releasing the next version.

Cheers,
Filip

PS: to be more specific, the error I'm getting is coming from 
platform/emulator/components.cc line 325, and looks like:
%***************** Error: distributed programming ***************
%**
%** Write failed during save
%**
%** File:  'some_file'
%** Error: 'No error'

("no error" because components.cc uses "errno" for reporting, but since 
the error comes from send(), it should be taking it from 
WSAGetLastError, which I'm sure reports 10038 "Not a socket handle").

Later, the same call returns:

%***************** Error: distributed programming ***************
%**
%** Write failed during save
%**
%** File:  'some_file'
%** Error: 'Permission denied'

Which I think is because some other calls in between set errno to 
"Permission denied".

So I think I'll fix the error reporting first and wait for your comments 
regarding hashtable implementation.

F.

> Dear Hackers,
>
> We had a discussion here at UCL about future releases of Mozart, and 
> we would
> like to have your opinion in order to take some decisions.
>
> First, there are two main projects that will produces releases 1.4.0 and
> 1.5.0. These are MozDSS and GeOz. The numbers will depend on which one
> releases first.
>
> But, since we haven't released since July 2006 (version 1.3.2), we 
> wonder if
> it is worthy to release 1.3.3 before MozDSS and-or GeOz. The point is 
> that no
> major bug fixes has been added to the 1-3-0-fixes branch, but here 
> there are
> some reasons in favour of doing the release:
>
> - There is now support for windoze Vista. This is a very small fix 
> that is
> already included in a new 1.3.2 installer for windowze, but it would 
> be good
> to uniform the packages.
> - The small fix in the oz script that calls the emulator, and that 
> concerns
> spaces in files names, seems to have fixed an annoying bug for some time.
> - The was an issue concerning the default behaviour of the OPI, where
> different threads where launched with every "declare" statement. This fix
> might be interesting mainly for students.
> - A 1.3.3 release would be a would opportunity to close the 
> 1-3-0-fixes branch
> - Releasing is good to show that the community is still working on 
> Mozart/Oz
> - Having a deadline for a release may force us to try to fix some 
> extra-small bugs
>
> The strongest argument against the release would be that the amount of 
> fixes
> is very small, and it might be enough to not release, but I'm not sure of
> that. This is why I would like to discuss this in order to take a 
> decision.
>
> Best,
> Boriss
>
>
>
>
> _________________________________________________________________________________ 
>
> mozart-hackers mailing list                           
> mozart-hackers at mozart-oz.org      
> http://www.mozart-oz.org/mailman/listinfo/mozart-hackers


-- 
Filip Konvička

LOGIS, s.r.o.
tel. +420 556 841 100
mob. +420 736 758 714