Re: Memory Barriers at pthread using CYGWIN

[Corinna Vinschen <corinna-cygwin@cygwin.com>]

On Jun 22 19:19, Takashi Yano via Cygwin wrote:
> Any suggestions?
> 
> On Tue, 20 Jun 2023 21:53:00 +0900
> Takashi Yano wrote:
> > I looked into this problem, and I think this is a problem regarding
> > _my_tls initialization order, so far. This seems to happen in LDAP
> > environment.
> > 
> > My assumption is:
> > 
> > If the program is the first program which load cygwin1.dll, ldap
> > connection seems to be made before pthread::init_mainthread().
> > In cyg_ldap.cc, cyg_ldap::connect(), cyg_ldap::search() or
> > cyg_ldap::next_page() calls cygwait() in which pthread::self()
> > is called.
> > 
> > Then, _my_tls.tid is initialized with null_pthread, therefore,
> > _my_tls.tid is not set in pthread::init_mainthread().
> > 
> > This causes pthread_join() failure at:
> > winsup/cygwin/thread.cc: line 2196
> >    if (!is_good_object (&joiner))
> >      return EINVAL;
> > 
> > 
> > The first idea to fix this issue is remove set_tls_self_pointer()
> > call from pthread::self().
> > 
> > diff --git a/winsup/cygwin/thread.cc b/winsup/cygwin/thread.cc
> > index 5c1284a93..a0f2d5546 100644
> > --- a/winsup/cygwin/thread.cc
> > +++ b/winsup/cygwin/thread.cc
> > @@ -392,10 +392,7 @@ pthread::self ()
> >  {
> >    pthread *thread = _my_tls.tid;
> >    if (!thread)
> > -    {
> > -      thread = pthread_null::get_null_pthread ();
> > -      thread->set_tls_self_pointer ();
> > -    }
> > +    thread = pthread_null::get_null_pthread ();
> >    return thread;
> >  }
> >  
> > 
> > The secnd approach is to re-initialize _my_tls.tid in
> > pthread::init_mainthread() if _my_tls.tid is null_pthread.
> > 
> > diff --git a/winsup/cygwin/thread.cc b/winsup/cygwin/thread.cc
> > index 5c1284a93..f614e01c4 100644
> > --- a/winsup/cygwin/thread.cc
> > +++ b/winsup/cygwin/thread.cc
> > @@ -364,7 +364,7 @@ void
> >  pthread::init_mainthread ()
> >  {
> >    pthread *thread = _my_tls.tid;
> > -  if (!thread)
> > +  if (!thread || thread == pthread_null::get_null_pthread ())
> >      {
> >        thread = new pthread ();
> >        if (!thread)
> > 
> > 
> > Which is the better approach, do you think?
> > Or any other idea?

The second approach looks good to me.  There was a reason to call
set_tls_self_pointer from pthread::self, I guess.  Resetting in
init_mainthread should have the least potential for side-effects.


Thanks,
Corinna