Re: [PATCH v2] locking/pvqspinlock: Add lock holder CPU argument to pv_wait()

[Waiman Long]

On 04/20/2016 10:36 AM, Pan Xinhui wrote:
> On 2016å04æ20æ 22:19, Peter Zijlstra wrote:
> > On Wed, Apr 20, 2016 at 10:15:09PM +0800, Pan Xinhui wrote:
> > > So there is such case that we search the whole hashtable and the lock is not found. :(
> > > Waiman assume that if l = null, the lock is not stored. however the lock might be there actually.
> > > But to avoid the worst case I just mentioned above, it can quickly finish the lookup.
> >
> > > > > +
> > > > > +			/*
> > > > > +			 * We try to locate the queue head pv_node by looking
> > > > > +			 * up the hash table. If it is not found, use the
> > > > > +			 * CPU in the previous node instead.
> > > > > +			 */
> > > > > +			hn = pv_lookup_hash(lock);
> > > > > +			if (!hn)
> > > > > +				hn = pn;
> > > > This is potentially expensive... it does not explain why this lookup can
> > > > fail etc.. nor mentioned that lock stealing caveat.
> > > Yes, it's expensive. Normally, PPC phyp don't always need the correct
> > > holder. That means current vcpu can just give up its slice.  There is
> > > one lpar hvcall H_CONFER. I paste some spec below.
> > Ok, so if we can indeed scan the _entire_ hashtable, then we really
> > should not have that in common code. That's seriously expensive.
> Okay, I will try to add the holder lookup code in arch/...
>
> But I just come up with one idea,
> in __pv_queued_spin_unlock_slowpath()
> we will kick the node->cpu, who will become the holder soon.
> I think we can somehow record the node->cpu and use it in pv_wait_node :)
>
> thanks
> xinhui

That will make the code more complex. Unless you find it provide real 
performance improvement, I won't suggest doing that for the time being.

Cheers,
Longman