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

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, 2 Feb 2019, Heiko Carstens wrote:

> On Sat, Feb 02, 2019 at 11:14:27AM +0100, Thomas Gleixner wrote:
> > On Sat, 2 Feb 2019, Heiko Carstens wrote:
> > So after the unlock @timestamp 337.215675 the kernel does not deal with
> > that futex at all until the failed lock attempt where it rightfully rejects
> > the attempt due to the alleged owner being gone.
> > 
> > So this looks more like user space doing something stupid...
> > 
> > As we talked about the missing barriers before, I just looked at
> > pthread_mutex_trylock() and that does still:
> > 
> > 	if (robust)
> >           {
> >             ENQUEUE_MUTEX_PI (mutex);
> >             THREAD_SETMEM (THREAD_SELF, robust_head.list_op_pending, NULL);
> >           }
> > 
> > So it's missing the barriers which pthread_mutex_lock() has. Grasping for
> > straws obviously....

Looks more like a solid tree than a straw now. :)

> Excellent! Taking a look into the disassembly of nptl/pthread_mutex_trylock.o
> reveals this part:
> 
> 140:   a5 1b 00 01             oill    %r1,1
> 144:   e5 48 a0 f0 00 00       mvghi   240(%r10),0   <--- THREAD_SETMEM (THREAD_SELF, robust_head.list_op_pending, NULL);
> 14a:   e3 10 a0 e0 00 24       stg     %r1,224(%r10) <--- last THREAD_SETMEM of ENQUEUE_MUTEX_PI

Awesome.

> I added a barrier between those two and now the code looks like this:
> 
> 140:   a5 1b 00 01             oill    %r1,1
> 144:   e3 10 a0 e0 00 24       stg     %r1,224(%r10)
> 14a:   e5 48 a0 f0 00 00       mvghi   240(%r10),0
> 
> Looks like this was a one instruction race...

Fun. JFYI, I said that I reversed the stores in glibc and on my x86 test VM
it took more than _3_ days to trigger. But the good news is, that the trace
looks exactly like the ones you provided. So it looks we are on the right
track.

> I'll try to reproduce with the patch below (sprinkling compiler
> barriers just like the other files have).

Looks about right.

Thanks,

	tglx