Re: Re: Re: Re: Re: Re: Re: [RFC PATCH 0/2] kvm/vmx: Output TSC offset

[Marcelo Tosatti <mtosatti@redhat.com>]

On Fri, Nov 30, 2012 at 10:36:43AM +0900, Yoshihiro YUNOMAE wrote:
> Hi Marcelo,
> 
> >>>Can you please write a succint but complete description of the method
> >>>so it can be verified?
> >>
> >>Sure.
> >>
> >>- Prerequisite
> >>1. the host TSC is synchronized and stable.
> >>2. kvm_write_tsc_offset events include previous and next TSC offset
> >>    values.
> >>3. Every event's trace_clock is TSC.
> >>
> >>- Assumption for the sake of ease
> >>1. One VCPU
> >>2. The guest executes no write_tsc or write_tsc only once.
> >>
> >>- Configuration
> >>1. On the host, kvm_exit/entry events are recorded in the buffer A and
> >>    kvm_write_tsc_offset events are recorded in the buffer B.
> >>2. Boot a guest
> >>
> >>- Sort method
> >>1.
> >>  Confirm which the kvm_write_tsc_offset events are recorded except for
> >>boot. Note that a vcpu thread writes TSC offset when boot as an initial
> >>operation.
> >>
> >>1-1.
> >>  If not recorded, it means that the guest did not execute write_tsc.
> >>So, we convert the guest TSC to the host TSC using TSC offset of boot.
> >>=> END
> >>
> >>1-2.
> >>  If recorded, it means that the guest executed write_tsc.
> >>So, we use new kvm_write_tsc_offset event information.
> >>
> >>2.
> >>We convert the host TSC(Th) of the kvm_write_tsc_offset event to
> >>the guest TSC(Tg) using previous_tsc_offset value:
> >>     Tg = Th + previous_tsc_offset
> >>
> >>3.
> >>To search the point where the guest executed write_tsc, we find "n"
> >>which satisfies the condition Tn < Tg < Tn+1 or Tn+1 < Tn < Tg from
> >>older events of the guest.
> >>The former condition means trace data of
> >>the guest were recorded monotonically. On the other hand, the latter
> >>condition means trace data of the guest moved backward.
> >>4.
> >>We convert the guest TSC of trace data to the host TSC using
> >>previous_tsc_offset value before "n" and using next_tsc_offset value
> >>after "n+1".
> >>=> END
> >>
> >>- Note
> >>We assumed one vcpu and no write_tsc or write_tsc only once for the
> >>sake of ease. For other conditions, we will use similar method.
> >>
> >>Thanks,
> >
> >There is no certainty. Consider the following information available
> >
> >guest trace			host trace
> >				100: guest_tsc_write (tsc_offset=-100 => guest_tsc = 0)
> >				104: guest_tsc_write (tsc_offset=-104 => guest_tsc = 0)
> >				108: guest_tsc_write (tsc_offset=-108 => guest_tsc = 0)
> >1: eventA 			
> >2: eventB
> >3: eventC
> >1: eventD
> >2: eventE
> >3: eventF
> >
> >How can you tell which tsc_offset to use for eventD ? It could be either
> >-104 or -108. The notion of "next_tsc_offset" is subject to such
> >issue.
> >
> >I suppose its fine to restrict the interface as follows: the tool will
> >accept one trace of events with monotonic timestamps and the user is
> >responsible for correlating that to a host trace.
> 
> OK, I'll add the restriction, which trace data must have monotonic
> timestamps to use this feature. I think guests seldom will execute
> write_tsc, so in many cases, timestamps will be monotonically recorded
> in trace data.
> 
> >That is, you can't feed distinct instances of guest kernel trace.
> 
> I'm not clear for "distinct instances". Is this about SMP or multiple
> guests? Would you explain about this?

Distinct boot instances. If the guest reboots TSC can be written to.