Re: [RFC][PATCH 4/4] futex: Rewrite FUTEX_UNLOCK_PI

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, 8 Oct 2016, Peter Zijlstra wrote:
> On Sat, Oct 08, 2016 at 05:53:49PM +0200, Thomas Gleixner wrote:
> > On Fri, 7 Oct 2016, Peter Zijlstra wrote:
> > > Solve all that by:
> > > 
> > >  - using futex specific rt_mutex calls that lack the fastpath, futexes
> > >    have their own fastpath anyway. This makes that
> > >    rt_mutex_futex_unlock() doesn't need to drop rt_mutex::wait_lock
> > >    and the unlock is guaranteed if we manage to update user state.
> > > 
> > >  - make futex_unlock_pi() drop hb->lock early and only use
> > >    rt_mutex::wait_lock to serialize against rt_mutex waiters
> > >    update the futex value and unlock.
> > > 
> > >  - in case futex and rt_mutex disagree on waiters, side with rt_mutex
> > >    and simply clear the user value. This works because either there
> > >    really are no waiters left, or futex_lock_pi() triggers the
> > >    lock-steal path and fixes up the WAITERS flag.
> > 
> > I stared at this for a few hours and while I'm not yet done analyzing all
> > possible combinations I found at least one thing which is broken:
> > 
> > CPU 0				CPU 1
> > 
> > unlock_pi(f)
> >   ....
> >   unlock(hb->lock)
> >   *f = new_owner_tid | WAITERS;
> > 
> > 				lock_pi(f) 
> > 				  lock(hb->lock)
> > 				  uval = *f;
> > 				  topwaiter = futex_top_waiter();
> > 				    attach_to_pi_state(uval, topwaiter->pistate);
> > 				      pid = uval & TID_MASK;
> > 				      if (pid != task_pid_vnr(pistate->owner))
> > 				      	 return -EINVAL;
> >   ....
> >   pistate->owner = newowner;
> > 
> > So in this case we tell the caller on CPU 1 that the futex is in
> > inconsistent state, because pistate->owner still points to the unlocking
> > task while the user space value alread shows the new owner. So this sanity
> > check triggers and we simply fail while we should not. It's [10] in the
> > state matrix above attach_to_pi_state().
> 
> Urgh, yes. I think I can cure that, by taking
> pi_state->pi_mutex.wait_lock in attach_to_pi_state(), but blergh.

There is another problem with all that racing against fixup_owner()
resp. fixup_pi_state_owner().

I fear, we need to rethink this whole locking/protection scheme from
scratch.

Thanks,

	tglx