Re: [PATCH 2/2] KVM: VMX: Extend VMX's #AC handding

[Andy Lutomirski]

> On Jan 31, 2020, at 12:18 PM, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> 
> ïOn Sat, Feb 01, 2020 at 01:47:10AM +0800, Xiaoyao Li wrote:
>>> On 1/31/2020 11:37 PM, Andy Lutomirski wrote:
>>> 
>>>> On Jan 30, 2020, at 11:22 PM, Xiaoyao Li <xiaoyao.li@xxxxxxxxx> wrote:
>>>> 
>>>> On 1/31/2020 1:16 AM, Andy Lutomirski wrote:
> 
> ...
> 
>>>>> Can we get a credible description of how this would work? I suggest: Intel
>>>>> adds and documents a new CPUID bit or core capability bit that means
>>>>> âsplit lock detection is forced onâ.  If this bit is set, the MSR bit
>>>>> controlling split lock detection is still writable, but split lock
>>>>> detection is on regardless of the value.  Operating systems are expected
>>>>> to set the bit to 1 to indicate to a hypervisor, if present, that they
>>>>> understand that split lock detection is on.  This would be an SDM-only
>>>>> change, but it would also be a commitment to certain behavior for future
>>>>> CPUs that donât implement split locks.
>>>> 
>>>> It sounds a PV solution for virtualization that it doesn't need to be
>>>> defined in Intel-SDM but in KVM document.
>>>> 
>>>> As you suggested, we can define new bit in KVM_CPUID_FEATURES (0x40000001)
>>>> as KVM_FEATURE_SLD_FORCED and reuse MSR_TEST_CTL or use a new virtualized
>>>> MSR for guest to tell hypervisor it understand split lock detection is
>>>> forced on.
>>> 
>>> Of course KVM can do this. But this missed the point. Intel added a new CPU
>>> feature, complete with an enumeration mechanism, that cannot be correctly
>>> used if a hypervisor is present.
>> 
>> Why it cannot be correctly used if a hypervisor is present? Because it needs
>> to disable split lock detection when running a vcpu for guest as this patch
>> wants to do?
> 
> Because SMT.  Unless vCPUs are pinned 1:1 with pCPUs, and the guest is
> given an accurate topology, disabling/enabling split-lock #AC may (or may
> not) also disable/enable split-lock #AC on a random vCPU in the guest.
> 
>>> As it stands, without specific hypervisor and guest support of a non-Intel
>>> interface, it is *impossible* to give architecturally correct behavior to a
>>> guest. If KVM implements your suggestion, *Windows* guests will still
>>> malfunction on Linux.
>> 
>> Actually, KVM don't need to implement my suggestion. It can just virtualize
>> and expose this feature (MSR_IA32_CORE_CAPABILITIES and MSR_TEST_CTRL) to
>> guest, (but it may have some requirement that HT is disabled and host is
>> sld_off) then guest can use it architecturally.
> 
> This is essentially what I proposed a while back.  KVM would allow enabling
> split-lock #AC in the guest if and only if SMT is disabled or the enable bit
> is per-thread, *or* the host is in "warn" mode (can live with split-lock #AC
> being randomly disabled/enabled) and userspace has communicated to KVM that
> it is pinning vCPUs.

How about covering the actual sensible case: host is set to fatal?  In this mode, the guest gets split lock detection whether it wants it or not. How do we communicate this to the guest?