Re: [PATCH v2] rwsem: Support optimistic spinning

[Tim Chen <tim.c.chen@linux.intel.com>]

On Wed, 2014-04-30 at 11:08 -0700, Tim Chen wrote:
> On Wed, 2014-04-30 at 20:04 +0200, Peter Zijlstra wrote:
> > On Wed, Apr 30, 2014 at 10:50:09AM -0700, Tim Chen wrote:
> > > On Wed, 2014-04-30 at 10:27 +0200, Peter Zijlstra wrote:
> > > > On Mon, Apr 28, 2014 at 03:09:01PM -0700, Davidlohr Bueso wrote:
> > > > > +/*
> > > > > + * Try to acquire write lock before the writer has been put on wait queue.
> > > > > + */
> > > > > +static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
> > > > > +{
> > > > > +	long count = ACCESS_ONCE(sem->count);
> > > > > +retry:
> > > > > +	if (count == RWSEM_WAITING_BIAS) {
> > > > > +		count = cmpxchg(&sem->count, RWSEM_WAITING_BIAS,
> > > > > +			    RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS);
> > 
> > count = RWSEM_WAITING_BIAS
> > new = RWSEM_WAITING_BIAS + RWSEM_ACTIVE_WRITE_BIAS
> > new = count + RWSEM_ACTIVE_WRITE_BIAS
> > 
> > > > > +		/* allow write lock stealing, try acquiring the write lock. */
> > > > > +		if (count == RWSEM_WAITING_BIAS)
> > > > > +			goto acquired;
> > > > > +		else if (count == 0)
> > > > > +			goto retry;
> > > > > +	} else if (count == 0) {
> > > > > +		count = cmpxchg(&sem->count, 0, RWSEM_ACTIVE_WRITE_BIAS);
> > 
> > count = 0
> > new = RWSEM_ACTIVE_WRITE_BIAS
> > new = count + RWSEM_ACTIVE_WRITE_BIAS
> > 
> > > > > +		if (count == 0)
> > > > > +			goto acquired;
> > > > > +		else if (count == RWSEM_WAITING_BIAS)
> > > > > +			goto retry;
> > > > > +	}
> > > > > +	return false;
> > > > > +
> > > > > +acquired:
> > > > > +	return true;
> > > > > +}
> > > > 
> > > > Could we have written that like:
> > > > 
> > > > static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
> > > > {
> > > > 	long old, count = ACCESS_ONCE(sem->count);
> > > > 
> > > > 	for (;;) {
> > > > 		if (!(count == 0 || count == RWSEM_WAITING_BIAS))
> > > > 			return false;
> > > > 
> > > > 		old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_BIAS);
> > > 
> > > Above line won't be correct for the case when count == 0.  We are trying 
> > > to acquire write lock, so the sem->count should become
> > > RWSEM_ACTIVE_WRITE_BIAS, or RWSEM_WAITING_BIAS + RWSEM_ACTIVE_BIAS.  
> > > So we should change the logic to
> > > 
> > > 		  if (count == RWSEM_WAITING_BIAS)
> > > 			old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_BIAS);
> > > 		  else
> > > 			old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_WRITE_BIAS);
> > 
> > I think I simply mis-typed it; shouldn't both cases be
> > RWSEM_ACTIVE_WRITE_BIAS ?
> 
> Yeah,  we should just write it as
> 			  old = cmpxchg(&sem->count, count, RWSEM_ACTIVE_WRITE_BIAS);

Oops, I mean
			    old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_WRITE_BIAS);

for count == 0, we need sem->count to be RWSEM_ACTIVE_WRITE_BIAS, 
for count == RWSEM_WAITING_BIAS, we need sem->count to be  RWSEM_WAITING_BIAS + RWSEM_ACTIVE_WRITE_BIAS

Tim