Wine, Windows.Forms on Linux, GC and segfaults.

[Miguel de Icaza]

Hello,

With the help of Hans Boehm, I have been tracking the problems we
have to run the Windows.Forms code with GC enabled.  Turns out that the
problem is not the Boehm code at all, it just exposes a problem that
might be happening elsewhere.

The problem is that by the time that Wine has been initialized,
using setjmp/longjmp will always lead to a crash.  The code in pthreads
that performs the longjmp will first try to invoke the pthread cleanup
routines, and then invoke longjmp.  This never happens.

In the particular case of Mono, I added a small bit of code before
calling into Boehm's GC, and then I run my app like this:

wine monostub.exe.so

The code snippet is:

jmp_buf buf;
    
    printf ("Here\n");
    if (setjmp (buf) == 0){
    	printf ("before\n");
    	longjmp (buf, 1);
    } else {
    	printf ("after\n");		
    }

The code should display:

Here
    before
    after

But with Wine, I get a crash inside longjmp, after the "before" is
printed out.  I get the Wine Console with the stack trace, but I can not
copy/paste from it, so I have included a screenshot of it.

I can not run Wine with Valgrind because it calls modify_ldt(),
but I suspect memory corruption.

I was wondering if it would be possible to just "link" against the
Wine library instead of having to create a stub application to run from?

Best wishes,
Miguel

[Steven Edwards]

>   I was wondering if it would be possible to just "link" against the
>
> Wine library instead of having to create a stub application to run from?

The way linking works in winelib is kinda odd with the usage of the *.spec that it seems this is
not possible. There was some disscussion on this a while back and certain changes that would be
needed to LD but I think those might also have been for PE support. If I remember right this has
something to do not only with LD but with alot of the magic wineserver works and hence just
linking your application to a winelib library would not include all of the support wineserver
provides. The rule seems to be its impossible to just -lwine your application. Wine is approaching
0.9 soon so once we hit 1.x then we should have a stable ABI/API which will make sharing the dlls
a lot easyer if you come up with a new loading/linking method for mono rather then the winebuild
system and its more interesting "features"

Thanks
Steven

__________________________________________________
Do you Yahoo!?
Yahoo! Mail Plus - Powerful. Affordable. Sign up now.
http://mailplus.yahoo.com

[Marcus Meissner]

On Tue, Dec 10, 2002 at 08:41:34PM -0500, Miguel de Icaza wrote:
> Hello,
> With the help of Hans Boehm, I have been tracking the problems we
>
> have to run the Windows.Forms code with GC enabled.  Turns out that the
> problem is not the Boehm code at all, it just exposes a problem that
> might be happening elsewhere.
> The problem is that by the time that Wine has been initialized,
>
> using setjmp/longjmp will always lead to a crash.  The code in pthreads
> that performs the longjmp will first try to invoke the pthread cleanup
> routines, and then invoke longjmp.  This never happens.
> In the particular case of Mono, I added a small bit of code before
>
> calling into Boehm's GC, and then I run my app like this:
>
> wine monostub.exe.so
>  The code snippet is:
>
> jmp_buf buf;
>
> printf ("Here\n");
>    if (setjmp (buf) == 0){
>    	printf ("before\n");
>    	longjmp (buf, 1);
>    } else {
>    	printf ("after\n");		
>    }
>  The code should display:
>
> Here
>    before
>    after
>   But with Wine, I get a crash inside longjmp, after the "before" is
>
> printed out.  I get the Wine Console with the stack trace, but I can not
> copy/paste from it, so I have included a screenshot of it.

You currently cannot use the wine libraries together with -lpthread.

WINE already overwrites and reimplements several pthread_ functions
which probably leads to internal confusion and the crash above.

Difficult to solve :/

Ciao, Marcus