Re: synchronize_qrcu_timeout()

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Wednesday, 27 June 2007 13:18, Oleg Nesterov wrote:
> On 06/26, Paul E. McKenney wrote:
> >
> > On Mon, Jun 25, 2007 at 07:49:57PM +0400, Oleg Nesterov wrote:
> > > On 06/25, Oleg Nesterov wrote:
> > > >
> > > > On 06/25, Paul E. McKenney wrote:
> > > > >
> > > > > On Mon, Jun 25, 2007 at 02:43:32PM +0400, Oleg Nesterov wrote:
> > > > > > 
> > > > > > Sadly, you can't use srcu/qrcu because it doesn't handle timeouts.	
> > > > > 
> > > > > Interesting...  So the thought is to have a synchronize_srcu_timeout()
> > > > > or something similar that waited for a grace period to elapse or for
> > > > > a timeout to expire, whichever comes first?  It should not be too hard
> > > > > to arrange something, if needed.
> > > > 
> > > > Yes. As for qrcu (see http://marc.info/?t=116484476500001), I think it is easy.
> > > > First, we add "int interrupted" into struct qrcu_struct, then something like this
> > > 
> > > Even simpler, we don't need ->interrupted.
> > 
> > I have to ask...
> > 
> > What sorts of performance characteristics are needed here?  The reason
> > that I ask is because putting a straight synchronize_qrcu() into a
> > workqueue (or something similar) and then using a timer to provide
> > any needed wakeup seems a lot simpler than rearranging the innards of
> > synchronize_qrcu().
> 
> Didn't think about that... But yes, we can implement synchronize_qrcu_timeout()
> on top of synchronize_qrcu() as you suggested. Hovewer, this implementation
> will add more code to .text compared to changing synchronize_qrcu().
> 
> Note also that we can't share the "context" of synchronize_qrcu_timeout() in
> that case. Each caller of synchronize_qrcu_timeout() needs a separate
> wait_queue_head_t + work_struct. This context can't live on stack, because
> it should survive after the timeout.
> 
> 
> If we change synchronize_qrcu() instead, we only add
> 
>  		if (unlikely(atomic_read(qp->ctr + prv))) {
>  			__wait_event_timeout(qp->wq, !atomic_read(qp->ctr + prv), tout);
>  			if (unlikely(!tout))
>  				goto out;
>  		}
> 
> to the slow path. This has nearly zero perfomance penalty. Yes, we have
> to wait if the previous call was interrupted by timeout. But in that case
> we lost nothing, we spend the same time waiting for qp->mutex when the
> previous call succeeds.
> 
> 
> That said, I am not sure synchronize_Xrcu_timeout() is terribly useful.
> Rafael could use it, but it is not better for this particular case.
> Except the code re-use, which is always good.

Well, if there are more cases in which it's useful, we'll be able to modify
this code to use it too in the future.

Greetings,
Rafael


-- 
"Premature optimization is the root of all evil." - Donald Knuth