Re: [PATCH v7 2/3] KVM: x86: trace mmio begin and complete

[Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>]

On 09/03/2012 07:07 PM, Avi Kivity wrote:
> On 08/27/2012 12:51 PM, Dong Hao wrote:
>> From: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
>>
>> 'perf kvm stat record/report' will use kvm_exit and kvm_mmio(read...) to
>> calculate mmio read emulated time for the old kernel, in order to trace
>> mmio read event more exactly, we add kvm_mmio_begin to trace the time when
>> mmio read begins, also, add kvm_io_done to trace the time when mmio/pio is
>> completed
>>
> 
> Why is this so critical?
> 
> If a lot of time is spent in in-kernel mmio, then 'perf top' will report
> it.  Otherwise the time between kvm_exit and kvm_entry describes the
> time spent in the host.  Not all of it is mmio handling, but it is quite
> close.

I have done some test, the new events can get more exact result, this is a
example in my slides:

the event handled time calculated by old events is: 0.000066(s)
calculated by new events, the result is: 0.000005(s).

So, i think it is worth introducing these new events.

But if you do not care it or think these events are duplicate with current
events or unmaintainable, i do not have strong opinion on that. Please let
me know, i will drop them in the next version.