[OLE #98] Re: By-pass RPC Runtime If Possible When Calling An STA

[Saulius Krasuckas]

* The log message from [1] says:
> By-pass the RPC runtime if possible when calling an STA by posting a
>    message directly to the apartment window for it to process. Fixes a
>    deadlock in InstallShield caused by having to create a thread when
>    freeing an object that comes from an STA apartment. Added tests that
>    fail without this fix.

After the change [2] running ole32 marshal test makes the process 
unkillable on my winME.  If I change the code, things get running OK:

|  static DWORD CALLBACK no_couninitialize_client_proc(LPVOID p)
|  {
|      struct ncu_params *ncu_params = (struct ncu_params *)p;
|      HRESULT hr;
|      IUnknown *pProxy = NULL;
|  
|      pCoInitializeEx(NULL, COINIT_APARTMENTTHREADED);
|  
|      hr = CoUnmarshalInterface(ncu_params->stream, &IID_IClassFactory, (void **)&pProxy);
|      ok_ole_success(hr, CoUnmarshalInterface);
|  
|      ok_more_than_one_lock();
|  
|      /* die without calling CoUninitialize */
| +    CoUninitialize();
|  
|      return 0;
|  }
|  
|  /* tests STA -> STA call during DLL_THREAD_DETACH doesn't deadlock */

It seems call deadlocks only on my machine.  Rob, can you tell me, please, 
how can I find out, is there some problem with my box setup or is this a 
bug in test-code?

[1] http://cvs.winehq.org/cvsweb/wine/dlls/ole32/tests/marshal.c
[2] http://cvs.winehq.org/cvsweb/wine/dlls/ole32/tests/marshal.c.diff?r1=1.24&am...

[Robert Shearman]

Saulius Krasuckas wrote:
> - The log message from [1] says:
> > By-pass the RPC runtime if possible when calling an STA by posting a
> >    message directly to the apartment window for it to process. Fixes a
> >    deadlock in InstallShield caused by having to create a thread when
> >    freeing an object that comes from an STA apartment. Added tests that
> >    fail without this fix.
>
> After the change [2] running ole32 marshal test makes the process 
> unkillable on my winME.

That test needs to be completely removed then. I'm surprised the Win9x 
implementation of COM differs so much from the WinNT version.

[Saulius Krasuckas]

* On Mon, 25 Jul 2005, Robert Shearman wrote:
> - Saulius Krasuckas wrote:
> > - The log message from [1] says:
> > > By-pass the RPC runtime if possible when calling an STA by posting a 
> > > message directly to the apartment window for it to process. Fixes a 
> > > deadlock in InstallShield caused by having to create a thread when 
> > > freeing an object that comes from an STA apartment. Added tests that 
> > > fail without this fix.
> >
> > After the change [2] running ole32 marshal test makes the process 
> > unkillable on my winME.
>
> That test needs to be completely removed then. I'm surprised the Win9x 
> implementation of COM differs so much from the WinNT version.

Wait a minute.  The funny thing is that the test didn't fail on Win95 and 
Win98 [3]!  It does on *my* WinME.  As I see no other WinME testers, I 
guess only my box' setup is at fault.

[3] http://test.winehq.org/data/200507211000/
    http://test.winehq.org/data/200507221000/

[Saulius Krasuckas]

* On Tue, 26 Jul 2005, Saulius Krasuckas wrote:
> - On Mon, 25 Jul 2005, Robert Shearman wrote:
> > - Saulius Krasuckas wrote:
> > > - The log message from [1] says:
> > > > By-pass the RPC runtime if possible when calling an STA by posting 
> > > > a message directly to the apartment window for it to process. 
> > > > Fixes a deadlock in InstallShield caused by having to create a 
> > > > thread when freeing an object that comes from an STA apartment. 
> > > > Added tests that fail without this fix.
> > >
> > > After the change [2] running ole32 marshal test makes the process 
> > > unkillable on my winME.
> >
> > That test needs to be completely removed then. I'm surprised the Win9x 
> > implementation of COM differs so much from the WinNT version.
>
> The funny thing is that the test didn't fail on Win95 and Win98 [3]!  
> It does on *my* WinME.  As I see no other WinME testers, I guess only my 
> box' setup is at fault.

I was right.  This happens due to the program, called FreeWheel [4].  It 
sets system-wide hook and plays with messages related to mouse wheel and 
scrolling.

It comes with no src AFAIK, and the only way to "fix" this is to quit 
FreeWheel.exe before I run winetest*.exe, I guess.

[4] http://www.geocities.com/SiliconValley/2060/freewheel.html