Re: [PATCH v2 0/4] KVM: x86: APICv cleanups

[Paolo Bonzini <pbonzini@redhat.com>]

On 22/10/21 16:56, Maxim Levitsky wrote:
>      vCPU0                                   vCPU1
>      =====                                   =====
> 
> - disable AVIC
>                                          - #NPT on AVIC MMIO access

vCPU1 is now OUTSIDE_GUEST_CODE

>                                          -*stuck on something prior to the page fault code*
> - enable AVIC
>                                          -*still stuck on something prior to the page fault code*
> - disable AVIC:
>    - raise KVM_REQ_APICV_UPDATE request

kvm_make_all_cpus_request does not wait for vCPU1

>    - zap the SPTE (does nothing, doesn't race
>         with anything either)

vCPU0 writes mmu_notifier_seq here

>                                          - now vCPU1 finally starts running the page fault code.

vCPU1 reads mmu_notifier_seq here

So yeah, I think you're right.

The VM value doesn't have this problem, because it is always stored 
before mmu_notifier_seq is incremented (on the write side) and loaded 
after the mmu_notifier_seq (on the page fault side).  Therefore, if 
vCPU1 sees a stale per-VM apicv_active flag, it is always going to see a 
stale mmu_notifier_seq.

With the per-vCPU flag, instead, the flag is written by 
kvm_vcpu_update_apicv, and that can be long after mmu_notifier_seq is 
incremented.

Paolo

>                                          - vCPU1 AVIC is still enabled
>                                            (because vCPU1 never handled KVM_REQ_APICV_UPDATE),
>                                            so the page fault code will populate the SPTE.
> 
>                                          - handle KVM_REQ_APICV_UPDATE
>                                             - finally disable vCPU1 AVIC
> 
>                                          - VMRUN (vCPU1 AVIC disabled, SPTE populated)