Re: [PATCH v6 06/13] KVM: nVMX: optimize prepare_vmcs02{,_full} for Enlightened VMCS case

[Paolo Bonzini]

On 17/10/2018 19:08, Jim Mattson wrote:
> I believe that ESXi reads GUEST_CS_AR_BYTES on every VM-exit to
> determine code size.

Which makes me wonder, maybe we should add GUEST_SS_AR_BYTES which is
where the CPL lives.  But then your tests from last year didn't find it.

Paolo

> On Wed, Oct 17, 2018 at 10:02 AM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>> On 17/10/2018 16:47, Vitaly Kuznetsov wrote:
>>>>> +   if (!hv_evmcs || !(hv_evmcs->hv_clean_fields &
>>>>> +                      HV_VMX_ENLIGHTENED_CLEAN_FIELD_GUEST_GRP2)) {
>>>>> +           vmcs_write16(GUEST_CS_SELECTOR, vmcs12->guest_cs_selector);
>>>>> +           vmcs_write32(GUEST_CS_LIMIT, vmcs12->guest_cs_limit);
>>>>> +           vmcs_write32(GUEST_CS_AR_BYTES, vmcs12->guest_cs_ar_bytes);
>>>>> +           vmcs_writel(GUEST_ES_BASE, vmcs12->guest_es_base);
>>>>> +           vmcs_writel(GUEST_CS_BASE, vmcs12->guest_cs_base);
>>>>> +   }
>>>> For what it's worth, I suspect that these can be moved to
>>>> prepare_vmcs02_full.  The initial implementation of shadow VMCS did not
>>>> expose "unrestricted guest" to the L1 hypervisor, and emulation does a
>>>> lot of accesses to CS (of course).  Not sure how ES base ended up in
>>>> there and not DS base, though...
>>> I tried unshadowing all these fields and at least Hyper-V on KVM
>>> (without using eVMCS of course) experiences a 1200-1300 cpu cycles
>>> regression during tight cpuid loop test. I checked and this happens
>>> because it likes vmreading GUEST_CS_AR_BYTES a lot.
>>
>> Go figure. :)  Liran, do you happen to know if ESX does something
>> similar with CS descriptor cache fields?
>>
>> Paolo