Re: WARN_ON_ONCE(!new_owner) within wake_futex_pi() triggerede

[Heiko Carstens <heiko.carstens@de.ibm.com>]

On Thu, Jan 31, 2019 at 01:27:25AM +0100, Thomas Gleixner wrote:
> On Thu, 31 Jan 2019, Thomas Gleixner wrote:
> 
> > On Wed, 30 Jan 2019, Paul E. McKenney wrote:
> > > On Thu, Jan 31, 2019 at 12:13:51AM +0100, Thomas Gleixner wrote:
> > > > I might be wrong as usual, but this would definitely explain the fail very
> > > > well.
> > > 
> > > On recent versions of GCC, the fix would be to put this between the two
> > > stores that need ordering:
> > > 
> > > 	__atomic_thread_fence(__ATOMIC_RELEASE);
> > > 
> > > I must defer to Heiko on whether s390 GCC might tear the stores.  My
> > > guess is "probably not".  ;-)
> > 
> > So I just checked the latest glibc code. It has:
> > 
> >             /* We must not enqueue the mutex before we have acquired it.
> >                Also see comments at ENQUEUE_MUTEX.  */
> >             __asm ("" ::: "memory");
> >             ENQUEUE_MUTEX_PI (mutex);
> >             /* We need to clear op_pending after we enqueue the mutex.  */
> >             __asm ("" ::: "memory");
> >             THREAD_SETMEM (THREAD_SELF, robust_head.list_op_pending, NULL);
> > 
> > 8f9450a0b7a9 ("Add compiler barriers around modifications of the robust mutex list.")
> > 
> > in the glibc repository, There since Dec 24 2016 ...
> 
> And of course, I'm using the latest greatest glibc for testing that, so I'm
> not at all surprised that it just does not reproduce on my tests.

As discussed on IRC: I used plain vanilla glibc version 2.28 for my
tests. This version already contains the commit you mentioned above.

> I just hacked the ordering and restarted the test. If the theory holds,
> then this should die sooner than later.

...nevertheless Stefan and I looked through the lovely disassembly of
_pthread_mutex_lock_full() to verify if the compiler barriers are
actually doing what they are supposed to do. The generated code
however does look correct.
So, it must be something different.