Re: [PATCH 2/6] kvm: x86: drop read_tsc_offset()

[Jim Mattson]

I'm thinking about the case where you want to merge traces from L0 and
L2. If the user code in L0 always knew the TSC offset of the current
VMCS, rather than the TSC offset of the L1 VMCS, this would be
trivial, regardless of the nature of L1.

On Wed, Sep 21, 2016 at 8:22 AM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>
>
> On 21/09/2016 17:19, Jim Mattson wrote:
>> Doesn't that assume you can run the merge program in L1?
>
> You only need the TSC offset, but we should make sure that L0
> tracepoints contain enough information to figure out the L0->L2 TSC
> offsets (they are the values in VMCS02).
>
> That said, how would you get the trace from L1 if you don't have access
> to it?
>
> Paolo
>
>> On Mon, Sep 19, 2016 at 10:37 PM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>>>
>>>
>>> On 20/09/2016 00:18, Jim Mattson wrote:
>>>> Hmmm. Yes, I think it does. With this patch series,
>>>> vcpu->arch.tsc_offset appears to contain L1's TSC offset (perhaps
>>>> making vmx->nested.vmcs01_tsc_offset redundant).
>>>>
>>>> However, this unfortunately limits the newly added functionality to
>>>> merging host and *L1* guest traces. It doesn't work with L2 (or
>>>> deeper) guests. Or perhaps I'm missing something?
>>>
>>> You can merge L1/L2 first and then host/L1.
>>>
>>> Paolo