Re: [PATCH] KVM: x86: Increase KVM_MAX_VCPUS to 4096

[Sean Christopherson <seanjc@google.com>]

On Wed, Sep 06, 2023, Vitaly Kuznetsov wrote:
> Sean Christopherson <seanjc@google.com> writes:
> 
> > On Tue, Aug 15, 2023, Kyle Meyer wrote:
> >> Increase KVM_MAX_VCPUS to 4096 when MAXSMP is enabled.
> >> 
> >> Notable changes (when MAXSMP is enabled):
> >> 
> >> * KMV_MAX_VCPUS will increase from 1024 to 4096.
> >> * KVM_MAX_VCPU_IDS will increase from 4096 to 16384.
> >> * KVM_HV_MAX_SPARSE_VCPU_SET_BITS will increase from 16 to 64.
> >> * CPUID[HYPERV_CPUID_IMPLEMENT_LIMITS (0x40000005)].EAX will now be 4096.
> >> 
> >> * struct kvm will increase from 39408 B to 39792 B.
> >> * struct kvm_ioapic will increase from 5240 B to 19064 B.
> >> 
> >> * The following (on-stack) bitmaps will increase from 128 B to 512 B:
> >> 	* dest_vcpu_bitmap in kvm_irq_delivery_to_apic.
> >> 	* vcpu_mask in kvm_hv_flush_tlb.
> >> 	* vcpu_bitmap in ioapic_write_indirect.
> >> 	* vp_bitmap in sparse_set_to_vcpu_mask.
> >> 
> >> Signed-off-by: Kyle Meyer <kyle.meyer@hpe.com>
> >> ---
> >> Virtual machines with 4096 virtual CPUs have been created on 32 socket
> >> Cascade Lake and Sapphire Rapids systems.
> >> 
> >> 4096 is the current maximum value because of the Hyper-V TLFS. See
> >> BUILD_BUG_ON in arch/x86/kvm/hyperv.c, commit 79661c3, and Vitaly's
> >> comment on https://lore.kernel.org/all/87r136shcc.fsf@redhat.com.
> >
> > Mostly out of curiosity, do you care about Hyper-V support?   If not, at some
> > point it'd probably be worth exploring a CONFIG_KVM_HYPERV option to allow
> > disabling KVM's Hyper-V support at compile time so that we're not bound by the
> > restrictions of the TLFS.
> >
> 
> (sorry for necroposting)

There's a big difference between mostly dead and all dead.  Mostly dead is slightly
alive. ;-)

> While adding CONFIG_KVM_HYPERV to disable all-things-Hyper-V may make
> sense for some deployments (and as we already have CONFIG_KVM_XEN), I
> don't think we should forbid KVM_MAX_VCPUS > 4096 when it is enabled:
> 'general purpose' (distro) kernels are used both for hosting large Linux
> guests and Windows guests. Instead, I'd suggest we define
> KVM_MAX_HV_VCPUS as MIN(KVM_MAX_VCPUS, 4096) and then e.g. fail
> KVM_SET_CPUID[,2] if we already have > 4096 vCPUs + fail
> kvm_arch_vcpu_create() if we already have something-hyperv enabled on
> the already created vCPUs.

Ah, bummer.  I'd still like to add CONFIG_KVM_HYPERV at some point, but I agree
it should be orthogonal to KVM_MAX_VCPUS.