lkml.org 
[lkml]   [2016]   [Oct]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH 4/4] futex: Rewrite FUTEX_UNLOCK_PI
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

\
 
 \ /
  Last update: 2016-10-10 12:20    [W:0.143 / U:1.488 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site