Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> On Fri, Feb 21, 2020 at 10:05:26PM +0800, linmiaohe wrote:
>> From: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>>
>> When kvm_vcpu_ioctl_get_cpuid2() fails, we set cpuid->nent to the
>> value of
>> vcpu->arch.cpuid_nent. But this is in vain as cpuid->nent is not
>> vcpu->copied to
>> userspace by copy_to_user() from call site. Get rid of this
>> meaningless assignment and further cleanup the var r and out jump label.
>
>Ha, took me a while to see that.
Sorry about it. I'am not good at it. :(
>>
>> On the other hand, when kvm_vcpu_ioctl_get_cpuid2() succeeds, we do
>> vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
>> - goto out;
>> - return 0;
>
>Hmm, so this ioctl() is straight up broken. cpuid->nent should be updated on success so that userspace knows how many entries were retrieved, i.e.
>the code should look something like below, with kvm_arch_vcpu_ioctl() unchanged.
>
>I'm guessing no VMM actually uses this ioctl(), e.g. neither Qemu or CrosVM use it, which is why the broken behavior has gone unnoticed. Don't suppose you'd want to write a selftest to hammer KVM_{SET,GET}_CPUID2?
>
>int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
> struct kvm_cpuid2 *cpuid,
> struct kvm_cpuid_entry2 __user *entries) {
> if (cpuid->nent < vcpu->arch.cpuid_nent)
> return -E2BIG;
>
> if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
> vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
> return -EFAULT;
>
> cpuid->nent = vcpu->arch.cpuid_nent;
>
> return 0;
>}
>
I searched KVM_GET_CPUID2 from Qemu, it's not used. So maybe we could just drop KVM_GET_CPUID2 altogether as
suggested by Paolo. Thanks for your review.
