Re: [PATCH 02/10] locking/qspinlock: Remove unbounded cmpxchg loop from locking slowpath

[Peter Zijlstra]

On Mon, Apr 09, 2018 at 06:19:59PM +0100, Will Deacon wrote:
> On Mon, Apr 09, 2018 at 05:54:20PM +0200, Peter Zijlstra wrote:
> > On Mon, Apr 09, 2018 at 03:54:09PM +0100, Will Deacon wrote:

> > > +/**
> > > + * set_pending_fetch_acquire - set the pending bit and return the old lock
> > > + *                             value with acquire semantics.
> > > + * @lock: Pointer to queued spinlock structure
> > > + *
> > > + * *,*,* -> *,1,*
> > > + */
> > > +static __always_inline u32 set_pending_fetch_acquire(struct qspinlock *lock)
> > > +{
> > > +       u32 val = xchg_relaxed(&lock->pending, 1) << _Q_PENDING_OFFSET;

	smp_mb();

> > > +       val |= (atomic_read_acquire(&lock->val) & ~_Q_PENDING_MASK);
> > > +       return val;
> > > +}

> > > @@ -289,18 +315,26 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
> > >  		return;
> > >  
> > >  	/*
> > > -	 * If we observe any contention; queue.
> > > +	 * If we observe queueing, then queue ourselves.
> > >  	 */
> > > -	if (val & ~_Q_LOCKED_MASK)
> > > +	if (val & _Q_TAIL_MASK)
> > >  		goto queue;
> > >  
> > >  	/*
> > > +	 * We didn't see any queueing, so have one more try at snatching
> > > +	 * the lock in case it became available whilst we were taking the
> > > +	 * slow path.
> > > +	 */
> > > +	if (queued_spin_trylock(lock))
> > > +		return;
> > > +
> > > +	/*
> > >  	 * trylock || pending
> > >  	 *
> > >  	 * 0,0,0 -> 0,0,1 ; trylock
> > >  	 * 0,0,1 -> 0,1,1 ; pending
> > >  	 */
> > > +	val = set_pending_fetch_acquire(lock);
> > >  	if (!(val & ~_Q_LOCKED_MASK)) {
> > 
> > So, if I remember that partial paper correctly, the atomc_read_acquire()
> > can see 'arbitrary' old values for everything except the pending byte,
> > which it just wrote and will fwd into our load, right?
> > 
> > But I think coherence requires the read to not be older than the one
> > observed by the trylock before (since it uses c-cas its acquire can be
> > elided).
> > 
> > I think this means we can miss a concurrent unlock vs the fetch_or. And
> > I think that's fine, if we still see the lock set we'll needlessly 'wait'
> > for it go become unlocked.
> 
> Ah, but there is a related case that doesn't work. If the lock becomes
> free just before we set pending, then another CPU can succeed on the
> fastpath. We'll then set pending, but the lockword we get back may still
> have the locked byte of 0, so two people end up holding the lock.
> 
> I think it's worth giving this a go with the added trylock, but I can't
> see a way to avoid the atomic_fetch_or at the moment.

So IIRC the addition of the smp_mb() above should ensure the @val load
is later than the @pending store.

Which makes the thing work again, right?

Now, obviously you don't actually want that on ARM64, but I can do that
on x86 just fine (our xchg() implies smp_mb() after all).


Another approach might be to use something like:

	val = xchg_relaxed(&lock->locked_pending, _Q_PENDING_VAL | _Q_LOCKED_VAL);
	val |= atomic_read_acquire(&lock->val) & _Q_TAIL_MASK;

combined with something like:

	/* 0,0,0 -> 0,1,1 - we won trylock */
	if (!(val & _Q_LOCKED_MASK)) {
		clear_pending(lock);
		return;
	}

	/* 0,0,1 -> 0,1,1 - we won pending */
	if (!(val & ~_Q_LOCKED_MASK)) {
		...
	}

	/* *,0,1 -> *,1,1 - we won pending, but there's queueing */
	if (!(val & _Q_PENDING_VAL))
		clear_pending(lock);

	...


Hmmm?