Re: [PATCH v2 01/49] KVM: x86: Do all post-set CPUID processing during vCPU creation

[Sean Christopherson]

On Thu, Jul 04, 2024, Maxim Levitsky wrote:
> On Fri, 2024-05-17 at 10:38 -0700, Sean Christopherson wrote:
> > diff --git a/arch/x86/kvm/cpuid.h b/arch/x86/kvm/cpuid.h
> > index 23dbb9eb277c..0a8b561b5434 100644
> > --- a/arch/x86/kvm/cpuid.h
> > +++ b/arch/x86/kvm/cpuid.h
> > @@ -11,6 +11,7 @@
> >  extern u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
> >  void kvm_set_cpu_caps(void);
> >  
> > +void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu);
> >  void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu);
> >  void kvm_update_pv_runtime(struct kvm_vcpu *vcpu);
> >  struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index d750546ec934..7adcf56bd45d 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -12234,6 +12234,7 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
> >  	kvm_xen_init_vcpu(vcpu);
> >  	kvm_vcpu_mtrr_init(vcpu);
> >  	vcpu_load(vcpu);
> > +	kvm_vcpu_after_set_cpuid(vcpu);
> 
> This makes me a bit nervous. At this point the vcpu->arch.cpuid_entries is
> NULL, but so is vcpu->arch.cpuid_nent so it sort of works but is one mistake
> away from crash.
>
> Maybe we should add some protection to this, e.g empty zero cpuid or
> something like that.

Hmm, a crash is actually a good thing.  In the post-KVM_SET_CPUID2 case, if KVM
accessed vcpu->arch.cpuid_entries without properly consulting cpuid_nent, the
resulting failure would be a out-of-bounds read.  Similarly, a zeroed CPUID array
would effectiely mask any bugs.

Given that KVM heavily relies on "vcpu" to be zero-allocated, and that changing
cpuid_nent during kvm_arch_vcpu_create() would be an extremely egregious bug,
a crash due to a NULL-pointer dereference should never escape developer testing,
let alone full release testing.

KVM does the "empty" array thing for IRQ routing (though in that case the array
and the nr_entries are in a single struct), and IMO it's been a huge net negative
because it's led to increased complexity just so that arch code can omit a NULL
check.