Re: [GnomeMeeting-devel-list] GnomeMeeting on 2.6, not working yet

[Damien Sandras <damien sandras it-optics com>]

I found the culprit.

GnomeMeeting is taking / releasing a mutex in a while loop which is
reading images from the videograbber. There was no sleep there because
it should be limited by the driver speed. That resulted, under the 2.6
kernel, to an extremely slow time before a call for the Wait method for
the same method returns in another thread. I suppose that it is also due
to the new scheduler and that the thread taking/releasing the mutex had
the priority, leading to a lower priority for the thread waiting for the
mutex to be released.

I added a very small sleep there and everything works fine for me now on
2.6 and on 2.4.

Kilian, time to update teh sources so that everybody can test and
report!

Thanks !
Le dim 01/02/2004 à 18:05, Damien Sandras a écrit :
> I did more testing, and under some circumstances (video transmission, no
> fast start, tunneling), GnomeMeeting takes extremely long to connect.
> 
> Unfortunately, it seems it is not the right fix yet :(
> 
> Le dim 01/02/2004 à 01:33, Christian Meder a écrit :
> > On Sat, 2004-01-31 at 16:00, Damien Sandras wrote:
> > > Hi all,
> > > 
> > > I can confirm it fixes the problem for me too!
> > > 
> > > Nice job!
> > > 
> > > I think Craig can commit the fix. Thanks very much!
> > > 
> > 
> > Hi,
> > 
> > I think we can scrap the last patch as it screws up gnomemeeting on
> > 2.4.x.
> > 
> > I nailed it down to the following patch against 20040131-CVS pwlib
> > version which works for me on 2.4.23 and 2.6.1.
> > 
> > I don't know why the ifdef Linux sched_yield was there in the first
> > place.
> > 
> > I'm very interested in other people's experiences with this patch.
> > Somebody still running 2.2 or 2.0 ?
> > 
> > 
> > 
> > 			Christian  
> > 
> > BTW nanosleep is a big difference between 2.4 and 2.6 because 2.6 is
> > running with HZ=1000 while 2.4 is HZ=100.
> > 
> > --- pwlib-cvs-20040131.CVS/src/ptlib/unix/tlibthrd.cxx-orig	2004-02-01 01:00:32.000000000 +0100
> > +++ pwlib-cvs-20040131.CVS/src/ptlib/unix/tlibthrd.cxx	2004-02-01 01:01:08.000000000 +0100
> > @@ -1053,25 +1053,21 @@
> >      pthread_testcancel();
> >  #endif
> >  
> >      lastTime = PTime();
> >    } while (lastTime < targetTime);
> >  }
> >  #endif
> >  
> >  void PThread::Yield()
> >  {
> > -  //sched_yield();
> > -  // re-add this if 2.4 becomes unusable..
> > -  // The following patch is improving 2.6 performance and it hopefully doesn't screw 2.4 ;)
> > -  struct timespec timer = { 0, 1 };
> > -  nanosleep (&timer, NULL);
> > +  sched_yield();
> >  }
> >  
> > 
> >  PThread * PThread::Current()
> >  {
> >    PProcess & process = PProcess::Current();
> >    process.threadMutex.Wait();
> >    PThread * thread = process.activeThreads.GetAt((unsigned)pthread_self());
> >    process.threadMutex.Signal();
> >    return thread;
> > @@ -1627,23 +1623,20 @@
> >    if (lockCount > 0) {
> >      lockCount--;
> >      return;
> >    }
> >  
> >    // otherwise mark mutex as available
> >    ownerThreadId = (pthread_t)-1;
> >  #endif
> >  
> >    PAssertPTHREAD(pthread_mutex_unlock, (&mutex));
> > -#ifdef P_LINUX
> > -  PThread::Yield();
> > -#endif
> >  }
> >  
> > 
> >  BOOL PMutex::WillBlock() const
> >  {
> >  #ifndef P_HAS_RECURSIVE_MUTEX
> >    pthread_t currentThreadId = pthread_self();
> >    if (currentThreadId == ownerThreadId)
> >      return FALSE;
> >  #endif
-- 
 _      Damien Sandras
(o-     
//\     It-Optics s.a.
v_/_    GnomeMeeting: http://www.gnomemeeting.org/
        FOSDEM 2004:  http://www.fosdem.org
        H.323 phone:  callto:ils.seconix.com/dsandras seconix com