Re: [RFC PATCH 2/2] kvm/arm64: expose hypercall_forwarding capability

[James Morse]

Hi Heyi,

On 24/09/2019 16:20, Heyi Guo wrote:
> Add new KVM capability "KVM_CAP_FORWARD_HYPERCALL" for user space to
> probe whether KVM supports forwarding hypercall.
> 
> The capability should be enabled by user space explicitly, for we
> don't want user space application to deal with unexpected hypercall
> exits. We also use an additional argument to pass exception bit mask,
> to request KVM to forward all hypercalls except the classes specified
> in the bit mask.
> 
> Currently only PSCI can be set as exception, so that we can still keep
> consistent with the old PSCI processing flow.

I agree this needs to be default-on, but I don't think this exclusion mechanism 
is extensible.


> diff --git a/arch/arm64/kvm/reset.c b/arch/arm64/kvm/reset.c
> index f4a8ae9..2201b62 100644
> --- a/arch/arm64/kvm/reset.c
> +++ b/arch/arm64/kvm/reset.c
> @@ -102,6 +105,28 @@ int kvm_arch_vm_ioctl_check_extension(struct kvm *kvm, 
> long ext)
>       return r;
>  }
>  
> +int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
> +                         struct kvm_enable_cap *cap)
> +{
> +     if (cap->flags)
> +             return -EINVAL;
> +
> +     switch (cap->cap) {
> +     case KVM_CAP_FORWARD_HYPERCALL: {
> +             __u64 exclude_flags = cap->args[0];

and if there are more than 64 things to exclude?


> +             /* Only support excluding PSCI right now */
> +             if (exclude_flags & ~KVM_CAP_FORWARD_HYPERCALL_EXCL_PSCI)
> +                     return -EINVAL;

Once we have a 65th bit, older kernels will let user-space set it, but nothing 
happens.


> +             kvm->arch.hypercall_forward = true;
> +             if (exclude_flags & KVM_CAP_FORWARD_HYPERCALL_EXCL_PSCI)
> +                     kvm->arch.hypercall_excl_psci = true;
> +             return 0;
> +     }
> +     }
> +
> +     return -EINVAL;
> +}

While 4*64 'named bits' for SMC/HVC ranges might be enough, it is tricky to 
work with.
Both the kernel and user-space have to maintain a list of bit->name and
name->call-number-range, which may change over time.

A case in point: According to PSCI's History (Section 7 of DEN022D), PSCIv1.1 
added
SYSTEM_RESET2, MEM_PROTECT and MEM_PROTECT_CHECK_RANGE.

I think its simpler for the HYPERCALL thing to act as a catch-all, and we 
provide
something to enumerate the list of function id's the kernel implements.

We can then add controls to disable the PSCI (which I think is the only one we 
have a case
for disabling). I think the PSCI disable should wait until it has a user.


Thanks,

James