[PATCH 1/5] KVM: arm64: Cap KVM_CAP_NR_VCPUS by KVM_CAP_MAX_VCPUS

[Marc Zyngier]

On Fri, 12 Nov 2021 09:51:10 +0000,
Vitaly Kuznetsov <vkuznets at redhat.com> wrote:
> 
> Marc Zyngier <maz at kernel.org> writes:
> 
> > Hi Vitaly,
> >
> > On 2021-11-11 16:27, Vitaly Kuznetsov wrote:
> >> It doesn't make sense to return the recommended maximum number of
> >> vCPUs which exceeds the maximum possible number of vCPUs.
> >> 
> >> Signed-off-by: Vitaly Kuznetsov <vkuznets at redhat.com>
> >> ---
> >>  arch/arm64/kvm/arm.c | 7 ++++++-
> >>  1 file changed, 6 insertions(+), 1 deletion(-)
> >> 
> >> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> >> index 7838e9fb693e..391dc7a921d5 100644
> >> --- a/arch/arm64/kvm/arm.c
> >> +++ b/arch/arm64/kvm/arm.c
> >> @@ -223,7 +223,12 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, 
> >> long ext)
> >>  		r = 1;
> >>  		break;
> >>  	case KVM_CAP_NR_VCPUS:
> >> -		r = num_online_cpus();
> >> +		if (kvm)
> >> +			r = min_t(unsigned int, num_online_cpus(),
> >> +				  kvm->arch.max_vcpus);
> >> +		else
> >> +			r = min_t(unsigned int, num_online_cpus(),
> >> +				  kvm_arm_default_max_vcpus());
> >>  		break;
> >>  	case KVM_CAP_MAX_VCPUS:
> >>  	case KVM_CAP_MAX_VCPU_ID:
> >
> > This looks odd. This means that depending on the phase userspace is
> > in while initialising the VM, KVM_CAP_NR_VCPUS can return one thing
> > or the other.
> >
> > For example, I create a VM on a 32 CPU system, NR_VCPUS says 32.
> > I create a GICv2 interrupt controller, it now says 8.
> >
> > That's a change in behaviour that is visible by userspace
> 
> Yes, I realize this is a userspace visible change. The reason I suggest
> it is that logically, it seems very odd that the maximum recommended
> number of vCPUs (KVM_CAP_NR_VCPUS) can be higher, than the maximum
> supported number of vCPUs (KVM_CAP_MAX_VCPUS).

I'm all for this change.

> All userspaces which use
> this information somehow should already contain some workaround for this
> case. (maybe it's a rare one and nobody hit it yet or maybe there are no
> userspaces using KVM_CAP_NR_VCPUS for anything besides complaining --
> like QEMU).
> 
> I'd like KVM to be consistent across architectures and have the same
> (similar) meaning for KVM_CAP_NR_VCPUS.

Sure, but this is a pretty useless piece of information anyway. As
Andrew pointed out, the information is available somewhere else, and
all we need to do is to cap it to the number of supported vcpus, which
is effectively a KVM limitation.

Also, we are talking about representing the architecture to userspace.
No amount of massaging is going to make an arm64 box look like an x86.

> > which I'm keen on avoiding. I'd rather have the kvm and !kvm cases
> > return the same thing.
> 
> Forgive me my (ARM?) ignorance but what would it be then? If we go for
> min(num_online_cpus(), kvm_arm_default_max_vcpus()) in both cases, cat
> this can still go above KVM_CAP_MAX_VCPUS after vGIC is created?

"min(num_online_cpus(), kvm_arm_default_max_vcpus())" is probably the
right thing in all cases. Yes, KVM_CAP_NR_VCPUS will keep reporting
more than the VM can actually support. But that's why we have
KVM_CAP_MAX_VCPUS, which tells you now many vcpus you can create for a
given configuration.

This shows how useless KVM_CAP_NR_VCPUS is, and I wouldn't mind a
documentation patch stating this.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.