2008/3/17, Bill Meier <wmeier@xxxxxxxxxxx>:
Peter Johansson wrote:
> 2008/3/13, Luis EG Ontanon <luis@xxxxxxxxxxx <mailto:luis@xxxxxxxxxxx>>:
>
> On Thu, Mar 13, 2008 at 2:38 PM, Stig Bjørlykke <stig@xxxxxxxxxxxxx
> <mailto:stig@xxxxxxxxxxxxx>> wrote:
> > I suppose we have to leave them allocated if we don't know how
> libsmi
> > is compiled?
> > I have almost no experience on Windows, do we compile this our self?
>
> I compiled it, I just made changes to the makefile (and build only the
> dll), I have no clue whether HAVE_VASPRINTF was defined or not.
>
>
> Did you compile it using MSVC 6? Could this then be of the same type of
> issue that is with adns, hence it is required to recompile libsmi too if
> using another compiler for Wireshark on Windows than MSVC 6. Developer
> Studio 2005EE for instance?
>
My 2 cents ....
1. As noted previously: the problem is that there are (in effect)
malloc(...) statements in libsmi which allocate storage and free(...)
statements in oids.c to free that storage.
This is a non-starter on Windows if libsmi.dll is linked
against a different msvcrt than wireshark. (Think allocating
memory in one heap and trying to free it in a completely
different heap). This is the case if Wireshark is built,
for example, using vc 2005EE(using msvcr80.dll) since
our version of the Windows libsmi is currently built
with vc6 (using msvcrt.dll).
(If this 'gotcha' isn't mentioned someplace in the developer docs, it
probably should be. If needed, I'll add something about the
whole issue of using mixed run-time MSVCRT libraries
with Wireshark on Windows).
2. Possible solutions:
a. (As a temporary workaround for anyone wishing to able to
build a working Wireshark with other than vc6 which uses
the current libsmi: Remove the free(...) statements in oids.c).
b. Compile libsmi as part of the build process (like adns & etc).
(It would be really nice not to have to do this ....).
c. Looking at the smi code: there's a smiFree function in
the smi file utils.c which seems like it might be
just what is needed.
However there's something to do with an
#ifdef HAVE_DMALLOC_H which I don't altogether understand. Maybe
this can be ignored since I'm pretty sure that our vanilla
Windows build done using the makefile mentioned previously
does not define this symbol.
If so, a solution might be (on Windows only):
1. exporting smiFree from libsmi (and then rebuilding
the library);
2. adding an extern smiFree(void *ptr) to oid.c;
3. changing the various free(...) statements in oids.c to
smiFree(...).
(Obviously it would be much cleaner if smiFree were available
as part of the API so that it could be used on any
platform. Or: maybe there's another better way .....
Possibly we could ask upstream if something can be done).
Also: I don't think it makes any difference whether
HAVE_VASNPRINTF & etc were defined when the Windows libsmi
was built.
Whether a library version of vasnprintf or the libsmi
version is used, storage is allocated and would need to be freed.
Thoughts ??
Bill
I believe that adding information to for instance README.developer on what is required to be done by developers on Windows due to the mix of msvcr80.dll and msvcrt.dll would be good. Would you (Bill) please add something in line with what you wrote above?
Personally I would want to go with 2c (for adns as well) since it is then possible to debug even this code. The pre-compiled libraries adns and smi will probably continue to generate problems since Windows based developers probably tend to shift more and more from VC6 to any of its successors (especially 2005EE).
With adns it should be easier than with smi since adns at least comes with build support for Windows. But there is no such thing for smi. I guess that's why it was added already compiled since nobody actually wants to maintain the makefiles of a third party library.
However, another approach might be possible here. One where nobody needs to maintain (at least not continuously) the makefiles for these libraries:
We could provide precompiled versions of both adns and smi using different versions of Visual Studio and Developer studio. It would then be possible to link the binaries with whatever version that matches the MSVC_VARIANT variable (config.nmake). Does anyone know if there are more versions of the MS VC runtime library than msvcr80.dll and msvcrt.dll? Is perhaps msvcrt.dll required only by VC6 and msvcr80.dll required for the rest of them?
/ Peter