Re: [PATCH v4] KVM: nVMX: Fully support of nested VMX preemption timer

[Paolo Bonzini <pbonzini@redhat.com>]

Il 16/09/2013 12:44, Gleb Natapov ha scritto:
> On Mon, Sep 16, 2013 at 11:49:40AM +0200, Paolo Bonzini wrote:
>> Il 16/09/2013 11:09, Gleb Natapov ha scritto:
>>> On Mon, Sep 16, 2013 at 10:58:12AM +0200, Paolo Bonzini wrote:
>>>>>>>  	vmx->__launched = vmx->loaded_vmcs->launched;
>>>>>>> +	if (is_guest_mode(vcpu) && !(vmx->nested.nested_vmx_exit))
>>>>>>> +		nested_adjust_preemption_timer(vcpu);
>>>>>>
>>>>>> Please leave the assignment to __launched last, since it's already
>>>>>> initializing the asm below.
>>>>>>
>>>>>> I don't like the is_guest_mode check here... Maybe it's
>>>>>> micro-optimizing, but I wonder if we already do too many checks in
>>>>>> vmx_vcpu_run...  For example, is_guest_mode could be changed (I think)
>>>>>> to a check for "vmx->loaded_vmcs == &vmx->vmcs1".
>>>>>>
>>>>> Why this will be more efficient that HF_GUEST_MASK check?
>>>>
>>>> Because we have already loaded vmx->loaded_vmcs, so it's one memory
>>>> access less.
>>>>
>>> But we will have to load vmx->vmcs1 instead :)
>>
>> That's not a memory load, it's an add.
>>
> You assume that vmx->loaded_vmcs and vmx will be both in registers here,
> isn't it too much to assume?

Both of them are needed in this very close line:

    vmx->__launched = vmx->loaded_vmcs->launched;

So it is not a big assumption---even on 32-bits, perhaps.  By contrast,
vcpu->hflags is very unlikely to be in a register.

This is of course mostly pointless alone.  But if we could shave 100
cycles off a lightweight vmexit, that could give a measurable
improvement on nested virt.

Paolo