Re: [PATCH v2 03/10] KVM: x86: Re-split x2APIC ICR into ICR+ICR2 for AMD (x2AVIC)

[Sean Christopherson <seanjc@google.com>]

On Thu, Aug 22, 2024, Tom Lendacky wrote:
> On 8/22/24 13:44, Sean Christopherson wrote:
> > +Tom
> > 
> > Can someone from AMD confirm that this is indeed the behavior, and that for AMD
> > CPUs, it's the architectural behavior?
> 
> In section "16.11 Accessing x2APIC Register" of APM Vol 2, there is this
> statement:
> 
> "For 64-bit x2APIC registers, the high-order bits (bits 63:32) are
> mapped to EDX[31:0]"
> 
> and in section "16.11.1 x2APIC Register Address Space" of APM Vol 2,
> there is this statement:
> 
> "The two 32-bit Interrupt Command Registers in APIC mode (MMIO offsets
> 300h and 310h) are merged into a single 64-bit x2APIC register at MSR
> address 830h."
> 
> So I believe this isn't necessary. @Suravee, agree?
> 
> Are you seeing a bug related to this?

Yep.  With APICv and x2APIC enabled, Intel CPUs use a single 64-bit value at
offset 300h for the backing storage.  This is what KVM currently implements,
e.g. when pulling state out of the vAPIC page for migration, and when emulating
a RDMSR(ICR).

With AVIC and x2APIC (a.k.a. x2AVIC enabled), Genoa uses the legacy MMIO offsets
for storage, at least AFAICT.  I.e. the single MSR at 830h is split into separate
32-bit values at 300h and 310h on WRMSR, and then reconstituted on RDMSR.

The APM doesn't actually clarify the layout of the backing storage, i.e. doesn't
explicitly say that the full 64-bit value is stored at 300h.  IIRC, Intel's SDM
doesn't say that either, i.e. KVM's behavior is fully based on throwing noodles
and seeing what sticks.

FWIW, AMD's behavior actually makes sense, at it provides a consistent layout
when switching between xAPIC and x2APIC.  Intel's does too, from the perspective
of being able to emit single loads/stores.