Re: [PATCH v2] x86/mm: Don't disable INVLPG if "incomplete Global INVLPG flushes" is fixed by microcode

[Andrew Cooper]

On 11/04/2024 6:38 am, Xi Ruoyao wrote:
> On Tue, 2024-04-09 at 08:56 +0100, Andrew Cooper wrote:
>> On 09/04/2024 2:43 am, Sean Christopherson wrote:
>>> On Mon, Apr 08, 2024, Michael Kelley wrote:
>>>> From: Dave Hansen <dave.hansen@xxxxxxxxx> Sent: Thursday, April 4, 2024 11:09 AM
>>>>> On 4/4/24 10:48, Michael Kelley wrote:
>>>>>> I agree one could argue that it is a hypervisor bug to present PCID to the guest
>>>>>> in this situation. It's a lot cleaner to not have a guest be checking FMS and
>>>>>> microcode versions. But whether that's practical in the real world, at least
>>>>>> for Hyper-V, I don't know. What's the real impact of running with PCID while
>>>>>> the flaw is still present? I don’t know the history here ...
>>>>> There's a chance that INVLPG will appear ineffective.
>>>>>
>>>>> The bad sequence would go something like this: The kernel does the
>>>>> INVLPG on a global mapping.  Later, when switching PCIDs, the TLB entry
>>>>> mysteriously reappears.  No PCIDs switching means no mysterious
>>>>> reappearance.
>>>> Xi Ruoyao's patch identifies these errata:  RPL042 and ADL063.  In the links
>>>> to the documents Xi provided, both of these errata have the following
>>>> statement in the Errata Details section:
>>>>
>>>>     This erratum does not apply in VMX non-root operation.  It applies only
>>>>     when PCIDs are enabled and either in VMX root operation or outside
>>>>     VMX operation.
>>>>
>>>> I don't have deep expertise on the terminology here, but this sounds
>>>> like it is saying the erratum doesn’t apply in a guest VM.  Or am I
>>>> misunderstanding?
>>> Huh.  My read of that is the same as yours.  If that's the case, then it probably
>>> makes sense to have KVM advertise support if PCID is available in hardware, even
>>> if PCID is disabled by the host kernel.
>> My reading is the same also.  Seems like VMs are fine.
> So... Should I sent a v6 with the hypervisor checking reverted [ i.e.
> always enable PCID if boot_cpu_has(X86_FEATURE_HYPERVISOR) ]?

If Linux can see a hypervisor, then yes it should be safe to use PCID.

~Andrew