[ECOS] Lightweight C++ multithreaded exceptions

[ECOS] Lightweight C++ multithreaded exceptions

[Øyvind Harboe]

One approach to supporting multithreaded exceptions that seem to work
quite well in my experience is to use the single threaded GCC toolchain,
but to make the the global variable in C++ exceptions, fc_static, per
thread. By 'per thread' I mean that just like a generic CPU register,
this global resource is part of the thread state.

The easiest way of doing this that came to mind, was to add code to the
eCos thread switching to save/restore the global variable on each thread
switch. This requires reusing a field in the thread class or adding a
new one. It was also suggested to make this support configureable via
CDL as it adds a *very* slight overhead to thread switching.

http://sources.redhat.com/ml/ecos-patches/2003-11/msg00047.html

However, it requires a patch to GCC to make fc_static visible (it is
currently declared static), which requires either get/set fn's or
changing the name to avoid polluting the global namespace, e.g.
_fc_static should do. 

Advantages:

- very easy to do. 2 lines of code, 30 lines of fluff(#if'defs', etc.)
  50 lines of documentation :-)
- GCC patch is simply changing:
	static void *fc_static;
	=>
	void *_fc_static;
- no need to distinguish between single threaded or multithreaded
  toolchains
- no need for multilib
- very efficient implementation
- no need for pthreads
- trivial change to eCos
- should work for all forseeable future

Disadvantages:

- requires a patch to GCC.



-- 
Øyvind Harboe
http://www.zylin.com


--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Lightweight C++ multithreaded exceptions

[Jerome Souquieres]

Øyvind Harboe wrote:

>One approach to supporting multithreaded exceptions that seem to work
>quite well in my experience is to use the single threaded GCC toolchain,
>but to make the the global variable in C++ exceptions, fc_static, per
>thread.
>

  That's interesting. However, from what I understand, this approach 
only solves the multithreaded exceptions problem. This does nothing to 
ensure that the standard C++ library (including STL) is thead safe. Am I 
right ?

--
  Jerome Souquieres


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Lightweight C++ multithreaded exceptions

[Øyvind Harboe]

On Mon, 2005-05-23 at 11:52 +0200, Jerome Souquieres wrote:
> 
>   That's interesting. However, from what I understand, this approach 
> only solves the multithreaded exceptions problem. This does nothing to
> ensure that the standard C++ library (including STL) is thead safe. Am
> I right ?

I know there are some atomicity issues, but is that already handled in a
CPU architecture specific but os-independent fashion by default in
libstdc++?

Does anyone know of a definitive list of issues w.r.t. libstdc++ &
multithreading?

I guess the fact that eCosCentric is destined to contribute their
libstdc++ support saps my initiative for contributing patches  in this
area. Even more so, because I have a solution for my needs.

-- 
Øyvind Harboe
http://www.zylin.com


--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Lightweight C++ multithreaded exceptions

[Jonathan Larmour]

Øyvind Harboe wrote:
> On Mon, 2005-05-23 at 11:52 +0200, Jerome Souquieres wrote:
> 
>>  That's interesting. However, from what I understand, this approach 
>>only solves the multithreaded exceptions problem. This does nothing to
>>ensure that the standard C++ library (including STL) is thead safe. Am
>>I right ?
> 
> 
> I know there are some atomicity issues, but is that already handled in a
> CPU architecture specific but os-independent fashion by default in
> libstdc++?

libstdc++ has atomicity primitives for some CPUs, including x86, m68k, 
mips, powerpc and SPARC. Notably it doesn't include ARM, probably because 
at least on some ARM architecture versions it isn't possible. I don't know 
for definite about recent ARM architecture versions.

But even that little atomicity support is only used in a handful of 
places, and on unsupported architectures it falls back to mutexes if 
available. However the key issue is that STL primarily uses OS mutexes if 
available, and also needs "once" variables. Without those, it is not 
thread-safe.

> Does anyone know of a definitive list of issues w.r.t. libstdc++ &
> multithreading?

There is support for multithreading, but you have to do it its way.

It's possible to hack it of course, but to do it the right way takes more 
effort. Certainly I had to make other changes to eCos (and the tools) to 
properly pass the libstdc++ testsuite. As you may have seen, there's a lot 
of tests there!

It's probably inappropriate for me to comment on your suggested changes to 
kernel rescheduling to support the fc_static munging, but personally I 
don't believe kernel modifications are the right way. I know, I know, 
sitting here with it all working, I would say that wouldn't I :-).

> I guess the fact that eCosCentric is destined to contribute their
> libstdc++ support saps my initiative for contributing patches  in this
> area. 

While that is ultimately true, as I indicated before, it will not be in 
the near future. Anyone who hopes to hang on a bit in their project for it 
to be released publically is likely to be disappointed.

 > Even more so, because I have a solution for my needs.

Your patches and instructions will certainly work in some situations. Some 
of it was using the code I included for our own C++ support anyway but did 
release :-), e.g. CYGPKG_LIBC_I18N_NEWLIB_CTYPE. But there are definitely 
a lot of gotchas once you introduce variation, including when you change 
to different architectures or try to use the libstdc++ library itself.

Jifl
-- 
--["No sense being pessimistic, it wouldn't work anyway"]-- Opinions==mine

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Lightweight C++ multithreaded exceptions

[Øyvind Harboe]

> Your patches and instructions will certainly work in some situations. Some 
> of it was using the code I included for our own C++ support anyway but did 
> release :-), e.g. CYGPKG_LIBC_I18N_NEWLIB_CTYPE. But there are definitely 
> a lot of gotchas once you introduce variation, including when you change 
> to different architectures or try to use the libstdc++ library itself.

I haven't taken my stuff through the libstdc++ testsuite. Mainly because
there is very little point. It does what I need and eCosCentric is
destined to release libstdc++ before these issues affect me. Or possibly
things will linger long enough that I make one more pass over the
libstdc++ stuff. No point in spending time until I know whether it is
necessary :-)

I've counted at least three projects that went from eCos to MMU-less
Linux variants when they tried to get STL to work(all had oodles of RAM,
which gives eCos less of an edge anyway) and I know there are at least a
handful of people using my hacked up stopgap solution. 


-- 
Øyvind Harboe
http://www.zylin.com


--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

RE: [ECOS] Lightweight C++ multithreaded exceptions

[Paul D. DeRocco]

> From: Jonathan Larmour
>
> libstdc++ has atomicity primitives for some CPUs, including x86, m68k,
> mips, powerpc and SPARC. Notably it doesn't include ARM, probably because
> at least on some ARM architecture versions it isn't possible. I
> don't know
> for definite about recent ARM architecture versions.

ARM7 and later have a swap byte or word, which is locked.

--

Ciao,               Paul D. DeRocco
Paul                mailto:pderocco@ix.netcom.com


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss