Re: sem_lock() vs qspinlocks

[Peter Zijlstra]

On Fri, May 20, 2016 at 05:05:05PM +0200, Peter Zijlstra wrote:
> On Fri, May 20, 2016 at 08:00:49AM -0700, Davidlohr Bueso wrote:
> > On Fri, 20 May 2016, Peter Zijlstra wrote:
> > 
> > >On Thu, May 19, 2016 at 10:39:26PM -0700, Davidlohr Bueso wrote:
> > >> In addition, this makes me wonder if queued_spin_is_locked() should then be:
> > >>
> > >>-	return atomic_read(&lock->val);
> > >>+	return atomic_read(&lock->val) & _Q_LOCKED_MASK;
> > >>
> > >>And avoid considering pending waiters as locked.
> > >
> > >Probably
> > 
> > Similarly, and I know you hate it, but afaict, then semantically
> > queued_spin_is_contended() ought to be:
> > 
> > -       return atomic_read(&lock->val) & ~_Q_LOCKED_MASK;
> > +       return atomic_read(&lock->val);
> 
> Nah, that would make it return true for (0,0,1), ie. uncontended locked.

FWIW, the only usage of spin_is_contended() should be for lock breaking,
see spin_needbreak().

This also means that

  #define spin_is_contended(l) (false)

is a valid implementation, where the only down-side is worse latency.

This is done (together with GENERIC_LOCKBREAK), to allow trivial
test-and-set spinlock implementations; as these cannot tell if the lock
is contended.