Re: [PATCH 2/5] kvmtrace: make cycle calculation architecture aware

["Yang, Sheng" <sheng.yang@intel.com>]

On Thursday 10 July 2008 21:32:29 Avi Kivity wrote:
> Yang, Sheng wrote:
> > On Wednesday 09 July 2008 23:03:19 Hollis Blanchard wrote:
> >> On Wed, 2008-07-09 at 11:17 +0200, Christian Ehrhardt wrote:
> >>> So the question that is left before changing that is, if the
> >>> original author had something special in mind chosing cycles
> >>> here. I added Eric on CC for that.
> >>>
> >>> I wait with my resubmission of the patch series until all
> >>> architectures agree *hope* on using getnstimeofday() - after an
> >>> ack from all sides I would revise my patch series and submit
> >>> that changes alltogether.
> >>
> >> I got an email bounce from Eric the last time I tried to email
> >> him, so I'm not sure he's still with Intel.
> >>
> >> However, I don't think he had any special intention; I think he
> >> was just porting xentrace to KVM.
> >
> > Eric had completed his internship in Intel, so...
> >
> > I like the term "timestamp" too. I think he used "cycles" only
> > because there is a function called get_cycles().
> >
> > But instead of getnstimeofday(), I suggest using ktime_get()
> > here. It's little more precise than getnstimeofday(), and ktime_t
> > is more easily to be handled. And I think the overhead it brought
> > can be ignored too.
>
> What is the overhead of ktime_get()?

Well, I just means it wrapped getnstimeofday(), and compared to 
rdtscll(), it got little overhead... :)

-- 
Thanks
Yang, Sheng