[PATCH v3 05/14] locking/qspinlock: Remove unbounded cmpxchg loop from locking slowpath

[will.deacon@arm.com (Will Deacon)]

On Sat, Apr 28, 2018 at 02:45:37PM +0200, Peter Zijlstra wrote:
> On Thu, Apr 26, 2018 at 05:55:19PM +0100, Will Deacon wrote:
> > On Thu, Apr 26, 2018 at 05:53:35PM +0200, Peter Zijlstra wrote:
> > > On Thu, Apr 26, 2018 at 11:34:19AM +0100, Will Deacon wrote:
> > > > @@ -290,58 +312,50 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
> > > >  	}
> > > >  
> > > >  	/*
> > > > +	 * If we observe any contention; queue.
> > > > +	 */
> > > > +	if (val & ~_Q_LOCKED_MASK)
> > > > +		goto queue;
> > > > +
> > > > +	/*
> > > >  	 * trylock || pending
> > > >  	 *
> > > >  	 * 0,0,0 -> 0,0,1 ; trylock
> > > >  	 * 0,0,1 -> 0,1,1 ; pending
> > > >  	 */
> > > > +	val = atomic_fetch_or_acquire(_Q_PENDING_VAL, &lock->val);
> > > > +	if (!(val & ~_Q_LOCKED_MASK)) {
> > > >  		/*
> > > > +		 * we're pending, wait for the owner to go away.
> > > > +		 *
> > > > +		 * *,1,1 -> *,1,0
> > > 
> > > Tail must be 0 here, right?
> > 
> > Not necessarily. If we're concurrently setting pending with another slowpath
> > locker, they could queue in the tail behind us, so we can't mess with those
> > upper bits.
> 
> Could be my brain just entirely stopped working; but I read that as:
> 
> 	!(val & ~0xFF) := !(val & 0xFFFFFF00)
> 
> which then pretty much mandates the top bits are empty, no?

Only if there isn't a concurrent locker. For example:


T0:
// fastpath fails to acquire the lock, returns val == _Q_LOCKED_VAL

if (val & ~_Q_LOCKED_MASK)
	goto queue;

// Fallthrough

T1:
// fastpath fails to acquire the lock, returns val == _Q_LOCKED_VAL

if (val & ~_Q_LOCKED_MASK)
	goto queue;

// Fallthrough

T0:
val = atomic_fetch_or_acquire(_Q_PENDING_VAL, &lock->val);
// val == _Q_LOCKED_VAL

T1:
val = atomic_fetch_or_acquire(_Q_PENDING_VAL, &lock->val);
// val == _Q_PENDING_VAL | _Q_LOCKED_VAL
// Queue into tail

T0:
// Spins for _Q_LOCKED_MASK to go to zero, but tail is *non-zero*


So it's really down to whether the state transitions in the comments refer
to the lockword in memory, or the local "val" variable. I think the former
is more instructive, because the whole thing is concurrent.

Will