Re: [PATCH 1/5] Add a global synchronization point for pvclock

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2010-04-19 at 13:47 +0300, Avi Kivity wrote:
> On 04/19/2010 01:43 PM, Peter Zijlstra wrote:
> >
> >>>
> >>>> +
> >>>>    cycle_t pvclock_clocksource_read(struct pvclock_vcpu_time_info *src)
> >>>>    {
> >>>>        struct pvclock_shadow_time shadow;
> >>>>        unsigned version;
> >>>>        cycle_t ret, offset;
> >>>> +    u64 last;
> >>>>
> >>>>
> >>>> +    do {
> >>>> +        last = last_value;
> >>>>          
> >>> Otherwise, this assignment can see a partial update.
> >>>        
> >> On a 32-bit guest, that is.
> >>      
> > Right, do bear in mind that the x86 implementation of atomic64_read() is
> > terrifyingly expensive, it is better to not do that read and simply use
> > the result of the cmpxchg.
> >
> >    
> 
> atomic64_read() _is_ cmpxchg64b.  Are you thinking of some clever 
> implementation for smp i386?


No, what I was suggesting was to rewrite that loop no to need the
initial read but use the cmpxchg result of the previous iteration.

Something like:

  u64 last = 0;

  /* more stuff */
 
  do {
    if (ret < last)
      return last;
    last = cmpxchg64(&last_value, last, ret);
  } while (last != ret);

That only has a single cmpxchg8 in there per loop instead of two
(avoiding the atomic64_read() one).